IJACSA Volume 16 Issue 11 - thesai.org

IJACSA Volume 16 Issue 11

Copyright Statement: This is an open access publication licensed under a Creative Commons Attribution 4.0 International License, which permits unrestricted use, distribution, and reproduction in any medium, even commercially as long as the original work is properly cited.

View Full Issue

Paper 1: Assessing LXC Containers on Raspberry Pi 4B/5 Boards in a Proxmox Virtual Environment

Abstract: On one hand, containerization is gaining acceptance as a lightweight virtualization alternative to Virtual Machines (VMs). On the other hand, Single-Board Computers (SBCs) are increasingly used due to their affordability, versatility, low energy consumption, and growing computational power. In this work, extensive experiments were conducted to assess the capabilities and limitations of Linux Containers (LXC) when deployed on a Proxmox Virtual Environment (Proxmox VE) cluster, built with Raspberry Pi (RPi) computers. The clusters consisted of either two Raspberry Pi 4 Model B (RPi 4B) or two Raspberry Pi 5 (RPi 5) with identical characteristics, connected through an Ethernet switch. The experiments aimed to determine: 1) the maximum number of containers that can be run simultaneously when varying their operating system, 2) the maximum number of containers that can be executed in parallel when varying their allocated RAM, 3) the maximum number of containers that can be run concurrently under different SBC memory configurations, 4) the time for container migration, and 5) the network performance between two containers. For storage, SATA SSDs were connected to the RPi 4B boards through their USB ports, while the RPi 5 boards used NVMe SSDs connected via their PCIe interfaces. The cluster formed with RPi 5 boards outperformed the one built with RPi 4B boards, showing significant improvements in the migration experiments. In terms of network performance, the results were similar between containers running on different nodes. However, much larger differences were observed between containers in execution on the same nodes. With this study, the authors aim to assist users and researchers in identifying and selecting the technologies and configurations that best meet the performance requirements of their specific study cases.

Author 1: Eric Gamess

Author 2: Deuntae Winston

Keywords: Container; virtualization; Proxmox VE; LXC; single-board computers; Raspberry Pi; performance evaluation

Paper 2: Agentic AI in Commodity Trading: A Comparative Simulation Study

Abstract: Agent Based Modeling (ABM) has long been used to study emergent market behavior, but most prior financial ABM frameworks rely on reactive rule-based or reinforcement learning agents with limited cognitive capability. This study introduces a novel integration of agentic artificial intelligence (AI) featuring autonomous goal setting, persistent memory, and multi-step planning into commodity trading simulations. We develop a hybrid ABM-Agentic AI framework and comparatively evaluate 20 traditional agents and 20 Agentic AI agents across Natural Gas and WTI Crude Oil markets over multiple horizons (1M–3Y). To address external validity concerns, synthetic price series are calibrated to historical volatility regimes. Results show consistent performance improvements for Agentic AI, with large practical effect sizes, although statistical significance is limited due to small sample sizes. We also identify sources of potential bias, such as higher initial skill ranges and frictionless execution, and present controlled adjustments to mitigate them. The study makes four contributions: 1) a novel simulation architecture for integrating cognitive AI into ABM; 2) explicit operationalization of agentic capabilities; 3) a controlled comparative evaluation across commodities; and 4) robustness checks examining sensitivity to volatility and parameter shifts. Limitations and recommendations for real data validation and realistic microstructure modeling are also discussed.

Author 1: TarakRam Nunna

Author 2: Ananya Samala

Keywords: Agentic AI; Agent Based Modeling; commodity trading; reinforcement learning; cognitive agents; market simulation

Paper 3: Analysis of Test Access Mechanisms for Improving Scan Compression and Test Time

Abstract: System-on-Chip (SoC) devices now integrate dozens—and sometimes hundreds—of heterogeneous embedded IP cores, each of which must be verified after fabrication. Industry therefore relies on modular testing so that every core can be exercised and validated without revealing its internal implementation, and so that designers can reuse test patterns efficiently. A persistent challenge is the mismatch between the limited scan-in/scan-out bandwidth at the chip boundary and a much larger channel capacity required if all cores were tested simultaneously. Widespread use of scan-compression schemes, such as Embedded Deterministic Test (EDT), offers several features for channel count selection, and this also needs to be considered during bandwidth allocation across cores. The multiple requirements are met using Test Access Mechanism (TAM), and over the past two decades, researchers have proposed many TAM architectures that move well beyond simple pin multiplexing, each balancing wiring overhead, concurrency, pattern compression, and scheduling complexity in different ways. However, a combined study of their effectiveness considering multiple aspects is not available. This study reviews the principles, algorithms, and architectures of TAM and test scheduling techniques. A classification of the techniques is provided, based on the method used and the area of application. The goal of the study is to create a platform for the future development of test access mechanisms. The study is believed to be helpful to both industry and academia.

Author 1: Vijay Sontakke

Keywords: Compression ratio; scan bandwidth; TAM; test coverage; test scheduling; test time

Paper 4: SUNDUS: A Human-Centered Framework for Fostering Human-AI Collaboration Through Transparency

Abstract: Imagine a future where a security operator can understand a complex threat in fractions of a second without specialized training. Not only that, but the operator can instantly understand the logic behind a particular AI warning. This vision of AI-augmented cognition is the focus of the proposed SUNDUS framework. However, today, this promise faces a critical barrier: the user. The immense potential of AI is often rendered useless when its operator lacks the professional training necessary to interpret and implement its outputs in the real world. This context reveals a critical research gap: while collaborative human-AI systems significantly enhance performance, their efficacy remains fundamentally dependent on extensive operator training, as evidenced by the OMAR (Operator Machine Augmentation Resource) system. The current paper proposes the SUNDUS (System for Understanding, Navigating, and Decision-Making Under Uncertainty and Support) framework, a theoretical model designed through a Human-Computer Interaction (HCI) lens to enhance human-AI collaboration by making system design a substitute for formal training. Leveraging principles from AMID (Augmented Multisensory Interface Design) and Visual Representations of Meta-Information, SUNDUS employs enhanced transparency—via Natural Language Explanations (NLEs), confidence scores, and multisensory cues—to offload cognitive burden and increase intuitive understanding. We propose a comparative experimental methodology to validate SUNDUS against OMAR, hypothesizing that SUNDUS will yield significantly higher decision-making accuracy and appropriately calibrated trust alongside a lower cognitive load in untrained users. The key implication is a scalable, human-centric design blueprint that shifts the burden of adaptation from the operator to the AI system, unlocking the full potential of augmented cognition.

Author 1: Abduljaleel Hosawi

Author 2: Richard Stone

Keywords: SUNDUS; human-AI collaboration; Human-Computer Interaction (HCI); transparency; multisensory display; cognitive load; decision-making; trust; training; OMAR

Paper 5: Leveraging Large Language Models in the Software Development Lifecycle: Opportunities and Challenges

Abstract: Large Language Models (LLMs) are increasingly integrated into software engineering workflows, yet existing studies provide fragmented or domain-specific examinations of their impact. This survey aims to systematically analyze how LLMs influence the Software Development Lifecycle (SDLC) end-to-end, identifying capabilities, limitations, risks, and emerging opportunities. We review 147 publications from 2017–2025 across ACM Digital Library, IEEE Xplore, ACL Anthology, and arXiv using predefined inclusion and exclusion criteria. Unlike prior surveys that focus narrowly on code generation or testing, this work provides an SDLC-wide synthesis supported by empirical benchmarks, industrial evidence, and a unified taxonomy mapping LLM capabilities to each phase of development. We further examine technical risks including hallucinations, dataset governance, robustness, security vulnerabilities, and auditability. The goal of this survey is to consolidate fragmented knowledge, highlight practical adoption challenges, and outline future research directions essential for building trustworthy, scalable, and effective LLM-enabled software engineering systems.

Author 1: Jasdeep Singh Bhalla

Author 2: Mansimar Kaur Jodhka

Keywords: Large Language Models (LLMs); Software Development Lifecycle (SDLC); AI-assisted software engineering; automated code generation; software testing; software architecture; DevOps automation

Paper 6: Evaluating Generalist Conversational AI Against Foundational Models of Instructional Design: A Comparative Analysis

Abstract: The rapid integration of Generative AI into instructional engineering presents a critical challenge: verifying the capacity of these tools to strictly adhere to systemic theoretical models of learning, despite the risk of generating "pedagogically hallucinated" content that possesses surface plausibility but lacks structural validity. This study addresses this gap by systematically evaluating the performance of generalist conversational AIs against foundational principles of Instructional Design (ID). Adopting a qualitative comparative analysis of four state-of-the-art models available in October 2025—GPT-5 (OpenAI), Gemini 2.5 Pro (Google), Claude Sonnet 4.5 (Anthropic), and DeepSeek V3.2 (DeepSeek AI)—we assessed their outputs for complex design scenarios against a multi-dimensional framework grounded in authoritative theories, including Biggs’s Constructive Alignment, Merrill’s First Principles of Instruction, and Universal Design for Learning (UDL). Results reveal a "paradox of competence without comprehension," where models demonstrate high factual reliability and linguistic fluency but exhibit significant shortcomings in maintaining logical pedagogical consistency, particularly regarding assessment alignment and accessibility standards, with only Claude Sonnet 4.5 demonstrating a notable proactive partnership posture. Consequently, we conclude that current generalist LLMs cannot function as autonomous expert designers and argue for a shift in professional practice toward Critical AI Literacy, where the human designer leverages AI for ideation but remains the essential guarantor of the pedagogical architecture.

Author 1: Abdelmounaim AZINDA

Author 2: Mohamed Khaldi

Keywords: Generative AI; large language models (LLMs); instructional design; constructive alignment; pedagogical evaluation; AI ethics in education; Human-AI collaboration

Paper 7: Anamel: Children’s Psychological and Mental Health Detection Application by Drawing Analysis Based on AI

Abstract: Psychological and mental health issues affect many people worldwide. However, the impact of these issues is stronger on children starting from early ages until their teenage. Using drawing to analyze and detect such feelings is a common way for specialists to help children express their feelings. Due to that, this study employs artificial intelligence to improve and ease this process by developing an application to help parents and specialists have a first-look analysis of their children's drawings. The Anamel application allows users to go through an experience that simulates clinics and psychological centers by being able to detect several feelings: happiness, sadness, anger, and aggression. Users start by answering pre-questions to collect initial information about the child. Then they can upload the image of the drawing, which is processed using computer vision techniques, where the AI model based on YOLO deep learning architecture provides the analysis results with an accuracy of 94%. Finally, they answer the post questions to ensure the final result. Specialists can also register themselves in the application to allow parents to communicate with them for extra help.

Author 1: Amal Alshahrani

Author 2: Manar Mohammed Almatrafi

Author 3: Jenan Ibrahim Mustafa

Author 4: Layan Saad Albaqami

Author 5: Raneem Abdulrahman Aljabri

Keywords: Application; mental health; drawings; artificial intelligence; YOLO; deep learning; computer vision; flutter

Paper 8: Enhanced Android Malware Detection Using Deep Learning and Ensemble Techniques

Abstract: Android malware continues to pose significant security threats, with evolving tactics that often bypass traditional detection systems. Existing detection mechanisms remain ineffective against obfuscated or novel malware variants, necessitating the development of more robust detection techniques. This study introduces a comprehensive machine learning framework for Android malware detection that leverages a systematic comparison between a deep Neural Network and diverse ensemble methods, including Voting Ensemble, Stacking Ensemble, XGBoost, and Random Forest. Unlike prior studies that often focus on individual approaches, this work provides an empirical benchmark that demonstrates how practical ensemble configurations can achieve superior performance while maintaining computational efficiency. The model is trained using the CIC-AndMal2017 dataset, incorporating a comprehensive set of static features, including API calls, permissions, services, receivers, and activities. Feature selection was performed to optimize model performance, reducing redundancy and improving detection accuracy. The models were evaluated on multiple classification metrics, including accuracy, F1-score, and confusion matrices, with the Voting Ensemble model achieving an accuracy of 94.14%, outperforming all other approaches, including the deep neural network. This study contributes to the field by demonstrating that a carefully constructed ensemble of diverse classifiers can not only improve detection accuracy but also offer a more scalable, lightweight solution compared to complex deep learning models. The research provides a significant advancement in practical Android malware detection by identifying optimal strategies that balance performance with computational efficiency.

Author 1: Abdul Museeb

Author 2: Yaman Hamed

Author 3: Rajalingam Sokkalingam

Author 4: Anis Amazigh Hamza

Author 5: Atta Ullah

Author 6: Iliyas Karim Khan

Keywords: Android malware detection; machine learning; API calls; permissions; android security; malware classification

Paper 9: Dual-Attention ResUNet-GAN for Secure Image Steganography: Optimizing the Trade-off Between Imperceptibility and Payload Capacity

Abstract: Secure and high-capacity data concealment has already become a requirement of modern multimedia communication, particularly with the enhanced protection and privacy levels of concern. The framework introduced in this study—the improved Dual-Attention ResUNet-GAN—helps optimize the trade-off among imperceptibility, robustness, and payload capacity in the field of image steganography. The two PatchGAN discriminators used in the model were a visual realism discriminator and a learned steganalyzer. Two encoders based on the ResNet-34 using CBAM-based dual attention are to be used. Just before the data is embedded, AES-256 encryption in CBC mode is employed to provide cryptographic confidentiality. Experiments on the COCO, BOSSbase, and ALASKA2 datasets are conducted to evaluate the proposed method's performance, yielding PSNR=42.5 dB, SSIM=0.98, BER=0.02, and high resistance to steganalysis (PE=91.2% vs. SRNet). Embedding is also changed in the proposed framework to high-entropy areas, thereby allowing the application of both conservative payloads (0.0156 bpp) and capacity-driven configurations (0.4 bpp) without affecting image quality. The findings have validated that the proposed system fits well with secure communication and intelligent data-hiding applications in real-world scenarios.

Author 1: Zobia Shabeer

Author 2: Muhammad Naeem

Author 3: Gohar Rahman

Author 4: Mehmood Ahmed

Author 5: Muhammad Zeeshan

Author 6: Asim Shahzad

Author 7: Salamah binti Fattah

Keywords: Image steganography; Generative Adversarial Networks (GANs); payload capacity; steganalysis robustness; artificial intelligence; BOSSbase; ALASKA#2

Paper 10: A Method for Planning the Dissemination Path of Traditional Chinese Medicine Culture Based on the Optimized Ant Colony Algorithm

Abstract: Strategic planning improves TCM cultural transmission efficacy, reliability, and impact. Many systems use heuristic or rule-based approaches, which have drawbacks such as path redundancy, low adaptation, and limited scalability in non-static networks. To address these constraints, we suggest RACO-TCM, or Reinforced Ant Colony Optimization for TCM Dissemination. This novel algorithmic distribution technique uses Ant Colony Optimization and reinforcement learning to create adaptable reward-driven cultural routes. The framework outperforms standard ant colony optimization because it uses dynamic pheromone updates, reinforcement-based exploration, and redundancy-aware heuristics to improve global search, convergence time, and robustness to local optimal solutions. We quantitatively assessed RACO-TCM against other methods and found that it increased cultural diffusion efficiency by 18.6% and reduced repeated routes by 12.3%. Creating a vast and instructive TCM knowledge graph with over 46,000 prescriptions, 8,000 herbs, and 25,000 chemical compounds achieved this. Overall, the TCM transmission technique is adaptive, scalable, and culturally consistent. It is used to manage business and TCM tourism, promote healthcare, digital education, and cultural services in smart cities.

Author 1: Qian Guo

Author 2: Ying Ma

Keywords: TCM dissemination; Ant Colony Optimization (ACO); intelligent path planning; cultural communication networks; knowledge graph optimization; algorithmic dissemination strategy

Paper 11: Maximizing Influence and Mitigating Harmful Viral Content in Social Networks

Abstract: This study presents an integrated framework containing sentiment and network analysis for social media modernization to mitigate the spread of harmful viral posts. The research focuses on detecting harmful content, identifying key influencers, and generating counter-narratives to promote constructive engagement. A Twitter Dataset was preprocessed to remove URLs, special characters, and numbers, and the VADER Tool was used to classify tweets into harmful, positive, and neutral categories. Network analysis was conducted by constructing directed retweet graphs to visualize information flow and identify influential users using the PageRank Algorithm and Vander centrality metrics. Counter-narratives were generated for harmful tweets to neutralize negativity and encourage positive discourse. Results show that integrating sentiment and network analysis reduces harmful content propagation by approximately 60 per cent through effective targeting of influential users. The proposed approach offers a scalable and data-driven model for social media moderation, contributing to safer and more ethical online communication environments by balancing freedom of expression with responsible content regulation.

Author 1: Muhammad Mohsin

Author 2: Muhammad Yaseen

Author 3: Umar Farooq Khattak

Author 4: Gohar Rahman

Keywords: Sentiment analysis; harmful content; counter-narratives; network analysis; PageRank; Vander influence maximization; digital ethics; social media moderation

Paper 12: Security Vulnerability Analysis and Enhancement of a Lightweight Sensor Node Authentication Framework

Abstract: This study presents a comprehensive structural and mathematical security analysis of LightAuth, a lightweight authentication framework, specifically designed for smart health sensor networks. We delve into its core components and identify several critical vulnerabilities that could compromise the integrity and security of the system. Our analysis reveals that the framework suffers from insufficient freshness verification, a flawed and biased key agreement process, and the persistent exposure of fixed identifiers, which makes it susceptible to various attacks. To address these significant security weaknesses, we propose a suite of practical and effective countermeasures. These enhancements include the implementation of a robust timestamp+nonce validation mechanism to ensure message freshness and the introduction of mutual signature verification to prevent man-in-the-middle attacks. Furthermore, we advocate for the use of dynamic pseudonyms to obfuscate user identities and enhance privacy. To bolster long-term security, we also integrate perfect forward secrecy (PFS), which ensures that a compromise of a long-term key does not compromise past session keys. We conducted extensive simulations to evaluate the effectiveness of these proposed enhancements. The results demonstrate that our improvements achieve a remarkable 100% replay detection rate, while the performance degradation remains within acceptable limits, proving the practicality of our solution.

Author 1: Kim Kyoung Yee

Author 2: Haewon Byeon

Keywords: Lightweight authentication; IoMT security; ECC; timestamp–nonce validation; replay resistance; formal verification; smart healthcare systems

Paper 13: A Proposed IoT System for Monitoring and Controlling Movable Zamzam Tanks in the Holy Mosque

Abstract: Zamzam water is very important for all Muslims worldwide, especially for Pilgrims and Umrah performers who are visiting the two holy mosques in Saudi Arabia. Many movable Zamzam tanks are dispersed in many places of the two holy mosques. These tanks are available for pilgrims and visitors for drinking. When the two holy mosques are crowded, visitors want to know the locations of the nearest filled Zamzam tanks. Administrators and operators also need to monitor tank statuses for performing necessary logistics operations. In this study, we proposed a novel IoT system for automating, monitoring, and controlling water levels in Zamzam tanks in the two holy mosques. The proposed system modifies the existing Zamzam tanks’ architecture to be smart tanks. The proposed system is analyzed, modeled, simulated, and discussed. Simulation and discussion show that the system is very useful for administrators, operators, and holy mosque visitors as it saves time, decreases efforts, automates logistics, and increases productivity. The system is very important for stakeholders, especially in crowded seasons. The performance metrics indicate that the proposed system is scalable with minimal delay and higher throughput. Also, the proposed system can participate in the digitalization of the two holy mosques, achieving the Saudi Vision 2030.

Author 1: Taha M. Mohamed

Keywords: Zamzam tanks monitoring; IoT smart systems; automatic monitoring and control; Logistics 5.0

Paper 14: Forecasting of Saudi Stock Prices Using Statistical and Machine Learning Models: A Multi-Model Comparative Approach

Abstract: Accurate forecasting of financial time-series data is not just a challenge—it's a critical necessity for investors in emerging markets. This study decisively evaluates the predictive power of seven advanced statistical and machine learning models: Autoregressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), Random Forest, eXtreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), K-Nearest Neighbors (KNN), and Decision Tree, across eight major stocks on the Saudi Stock Exchange (Tadawul). Employing a robust, lag-based forecasting framework, we meticulously assessed model performance using RMSE, R², directional accuracy, and computational efficiency. We introduce a hybrid evaluation framework that integrates magnitude accuracy, directional precision, and runtime profiling to guide model selection at the individual stock level, an approach that has not been previously applied to the Saudi market. The empirical evidence is compelling: model selection is heavily stock-specific. The classical ARIMA model consistently outperformed the others, delivering the lowest error and highest goodness-of-fit for stable, high-capitalization stocks, underscoring the timeless relevance of linear autoregressive components. Conversely, the ensemble method XGBoost emerged as a powerhouse of computational efficiency and predictive balance for more volatile series, boasting an optimal operational profile (runtime of ~ 1.5 s). While deep learning (LSTM) and SVR models fall short of magnitude metrics owing to the low signal-to-noise ratio in daily close price data, these findings offer practical guidance for investors, analysts, and policymakers seeking scalable, stock-specific forecasting strategies. Considering Saudi Arabia’s Vision 2030 and the increasing demand for real-time financial intelligence, this research addresses the urgent need for scalable stock-specific forecasting frameworks that support investor decision-making and policy formulation.

Author 1: Eissa Alreshidi

Keywords: Financial time-series forecasting; Saudi Stock Market; ARIMA; XGBoost; deep learning; ensemble models; LSTM; stock price prediction

Paper 15: Speech Emotion Recognition via Parallel Dual-Branch Fusion Model

Abstract: Speech Emotion Recognition (SER) has become a pivotal topic within affective computing and human–computer interaction, where the core challenge lies in jointly capturing both the time–frequency structure and the semantic context of speech. To overcome the shortcomings of current approaches—including single-view feature representation, the lack of emotional discriminability in self-supervised models, and suboptimal complementarity among fusion strategies—this study proposes a parallel dual-branch fusion architecture for SER. The framework consists of a wav2vec 2.0 branch and a CNN–Transformer spectrogram branch, which respectively extract contextual semantic representations from raw waveforms and explicit time–frequency features from spectrograms. A logistic regression fusion mechanism is further introduced at the decision level to achieve adaptive weighting in the probability space, thereby fully leveraging the complementary strengths of the two feature types. Experiments carried out on the RAVDESS audio subset showed that the proposed model surpassed several mainstream baselines (e.g., CNN-n-GRU and RELUEM), achieving 92.7% accuracy and 92.2% Macro-F1, with an average improvement of about 3.2 percentage points. The layer unfreezing studies confirmed the effectiveness of partial fine-tuning for transferring pretrained features, while the comparative experiments on fusion strategies validated the superiority of probability-space fusion in both performance and stability. Overall, the proposed framework achieves simultaneous gains in accuracy and robustness through feature complementarity, branch decoupling, and lightweight fusion. Future work will explore cross-lingual generalization, multimodal extensions, lightweight deployment, and dynamic emotion modeling, contributing to more efficient affective computing and intelligent interaction systems.

Author 1: Zhongliang Wei

Author 2: Chang Ge

Author 3: Lijun Zhu

Author 4: Jinmin Ye

Keywords: RAVDESS; Speech Emotion Recognition; spectrogram modeling; probability-space fusion; wav2vec 2.0 fine-tuning

Paper 16: An Improved Method Based on YOLOv7 for Detecting the Safety Helmets of Two-Wheeled Bicycle Riders

Abstract: Convolutional neural networks (CNNs) were widely used in object detection tasks. Usually, CNNs with strong object detection performance were difficult to apply to small, mobile embedded systems with limited computational resources due to the large number of parameters. Aiming at this problem, the lightweight improvement method for the safety helmet object detection task based on YOLOv7 has been studied. The first step was the lightweight improvement of the network. Taking YOLOv7 and YOLOv7-Tiny as the basic networks, respectively, the backbone network was improved using the MobileOne network. YOLOv7-MobileOne (YOLOv7-MO) and YOLOv7-Tiny-MobileOne (YOLOv7-TMO) were obtained. Compared with the original network parameters, the number of parameters decreased by 36.8% and 37.9%, respectively. Verified on the Pascal VOC dataset, the YOLOv7-MO had a 3.7% decrease in mAP @.5 compared to the YOLOv7. The YOLOv7-MO had a 9.8% increase in mAP @.5 compared to the YOLOv7-TMO. The second step was to improve the detection accuracy. The Coordinate Attention (CA) module was integrated at different positions of YOLOv7-MO and YOLOv7-TMO, respectively, to obtain YOLOv7-MO-Coordinate Attention (YOLOv7-MOC) and YOLOv7-TMO-Coordinate Attention (YOLOv7-TMOC). Verified on the Pascal VOC dataset, YOLOv7-MOC improved 1.44% compared to YOLOv7-MO's mAP @.5 and reduced FPS by 5.4Hz. Verified on the self-constructed two-wheeled cyclists helmet dataset (TCHD), YOLOv7-MOC increased by 0.8% compared to YOLOv7-MO's mAP @.5 and reduced FPS by 0.3Hz. YOLOv7-MOC increased by 1.0% compared to YOLOv7's mAP @.5 to 77.1%. The corresponding FPS was 28.7Hz higher, reaching 89.3Hz. Finally, experiments were conducted using the Raspberry Pi 4B embedded development board, based on the Linux system and the Pytorch framework, with the YOLOv7-TMOC network model. The results proved that the improved network model can be applied to the object detection of small embedded systems.

Author 1: Xufei Wang

Author 2: Penghui Wang

Author 3: Zishuo Wang

Author 4: Jeonyoung Song

Author 5: Jinde Song

Keywords: Object detection; YOLOv7; MobileOne; CA module; TCHD

Paper 17: Lightweight Multi-Feature Fusion GAN with Deformable Attention for HMD-Occluded Face Reconstruction

Abstract: Head-mounted displays (HMDs) enhance virtual reality (VR) experiences, but occlude the upper face, hindering realistic user representation. To address this, some studies employ sensors to capture facial expressions under occlusion, while deep learning methods typically rely on image inpainting to restore missing regions. However, these approaches often suffer from limitations such as insufficient shallow feature representation, high computational complexity, and redundant model structures. This study proposes a lightweight generative adversarial network (GAN) that utilizes multi-feature fusion and deformable attention for face reconstruction under HMD occlusion. Specifically, a Lie group feature learning module is used to enhance shallow geometric representations, while reference-guided deformable attention dynamically focuses on occluded regions, improving both structural fidelity and efficiency. Experiments across multiple face datasets show that the proposed method outperforms existing mainstream approaches regarding structural fidelity, detail restoration capability, and model efficiency. The proposed framework offers a promising solution for integration with HMDs equipped with facial tracking, enabling more realistic and expressive avatars in VR applications.

Author 1: Yingying Li

Author 2: Ajune Wanis Ismail

Author 3: Muhammad Anwar Ahmad

Author 4: Norhaida Mohd Suaib

Author 5: Fazliaty Edora Fadzli

Keywords: Generative adversarial network; Lie group feature learning; deformable attention; face reconstruction; virtual reality; head-mounted displays

Paper 18: Adult-Size Bio-Mechatronic Exoskeleton with Robotic Underactuated Mechanism for Continuous Passive Motion Therapy on Hemiplegic Lower Limbs

Abstract: Cerebrovascular accidents (CVA) cause hemiplegia in adult patients, limiting their daily activities and reducing their autonomy, which calls for effective rehabilitation systems for the lower limbs. The study aims to design and simulate a subactuated robotic continuous passive motion system for lower-limb therapy in adult patients with post-stroke hemiplegia. Physical rehabilitation protocols were analyzed, and consultations with specialists were conducted to define criteria related to safety, stability, and adaptability. The mechanical subsystem was developed using SolidWorks with an adjustable structure; the electronic subsystem was designed using Autodesk Eagle and KiCad 7 to enable precise control; and the software subsystem was implemented in Wokwi using an ESP-32 microcontroller for parameter configuration and data transmission. Simulations were conducted in CoppeliaSim, modeling a virtual patient undergoing three rehabilitation modes: knee flexion-extension, ankle flexion-extension, and a combined full knee extension cycle. Simulations validated the correct execution of the planned movements, stability of the power supply, precision in displacement control, and the ability to record and transmit therapeutic data. The strength of materials and the effectiveness of the mechanism were also confirmed. The designed subactuated robotic system represents a safe and effective tool for physical rehabilitation, with the potential to improve mobility and quality of life in patients with lower-limb hemiplegia following a CVA.

Author 1: Adrian Nacarino

Author 2: Carol Sandoval

Author 3: Cesar Martel

Author 4: Jose Cornejo

Author 5: Margarita Murillo

Author 6: Jeanette Borja

Author 7: Josue Alata Rey

Author 8: Ricardo Palomares

Keywords: Bio-mechatronics; engineering design; exoskeleton; stroke rehabilitation; rehabilitation robotics

Paper 19: Improving Spam Detection with Feature Engineering and Adaptive Learning Approaches

Abstract: Spam email detection is a critical component of securing and maintaining reliable digital communication systems. This study explores the effectiveness of various machine learning algorithms in classifying spam, with an emphasis on enhancing accuracy and precision through systematic preprocessing, advanced feature engineering, and text preprocessing. Six models were evaluated: Logistic Regression, Support Vector Classifier, Multinomial Naïve Bayes, K-Nearest Neighbors, AdaBoost, and Bagging Classifier using a comprehensive preprocessing pipeline that included Term Frequency–Inverse Document Frequency vectorization, feature scaling, and the incorporation of engineered features such as character counts. Experimental results reveal that Multinomial Naïve Bayes consistently achieved the highest precision 1.00 and strong accuracy 0.979 when paired with feature scaling, while Logistic Regression delivered robust and stable performance across multiple configurations with precision exceeding 0.96, making it a reliable choice for real-world deployment. Although Support Vector Classifier and AdaBoost exhibited competitive baseline performance, Support Vector Classifier showed limitations when handling numeric features, whereas AdaBoost maintained consistent results across scenarios. These findings underscore the critical role of tailored preprocessing and ensemble learning in improving classification outcomes and highlight the comparative strengths of different algorithms in real-world spam detection. In particular, Multinomial Naïve Bayes proved highly effective for precision-critical tasks, while Logistic Regression emerged as a dependable solution for environments requiring consistent reliability. Overall, this work advances machine learning-based spam filtering by identifying models that successfully balance precision, adaptability, and computational efficiency.

Author 1: Sadeem H. AlHomidan

Author 2: Marwah M. Almasri

Author 3: Shimaa A. Nagro

Keywords: Spam detection; machine learning; Multinomial Naïve Bayes; logistic regression; ensemble learning; text preprocessing; feature engineering

Paper 20: A Newton-Raphson-Based Optimizer-Driven Temporal Convolutional Networks for Birth Rate Prediction in a Small Area

Abstract: For economically developed small geographic regions, population forecasting serves as a vital tool for achieving refined regional management. However, due to relying on the subjective experience of experts, traditional methods for predicting birth rates have shortcomings in accuracy, resulting in unreliable results. To address this limitation, this study introduces deep learning (DL) models into the domain of birth rate prediction. Specifically, a hybrid TCN-Bi-LSTM model is proposed, integrating a Temporal Convolutional Network (TCN) with a Bi-directional Long Short-Term Memory (Bi-LSTM) network to predict birth populations in small regions. The proposed hybrid model effectively leverages the strengths of the TCN and Bi-LSTM to capture both local temporal patterns and long-term hidden dependencies within birth rate time series data. The proposed birth rate prediction model not only incorporates historical data on regional birth rates but also accounts for the influence of factors such as divorce rates, consumption levels, and population size. Furthermore, an enhanced meta-heuristic algorithm is designed to optimize the hyperparameters of the hybrid TCN-Bi-LSTM model, with the aim of increasing its prediction accuracy. The hippopotamus position update strategy was introduced into the Newton-Raphson-Based Optimizer (NRBO), and an improved NRBO (INRBO) algorithm was developed. Finally, the performance of the proposed birth rate prediction model was validated using a dataset from three regions or countries. The prediction results demonstrate that, compared to the other four models, the proposed INRBO-TCN–Bi-LSTM model achieves the best performance, with an average reduction of 95% in training loss.

Author 1: Shengyi Zhou

Author 2: Liang Chen

Author 3: Wei Han

Author 4: Bin Liu

Keywords: Temporal Convolutional Network; Bi-directional Long Short-Term Memory; prediction model; birth rate; meta-heuristic algorithm

Paper 21: A New Advanced Multimodal Sentiment Classification Through Combined Attention Mechanisms: A-MSDA

Abstract: Interest in multimodal sentiment analysis has grown significantly due to the widespread sharing of text and images on social platforms. Existing approaches often emphasize either sentiment features within textual–visual data or the correlation between modalities, leaving gaps in effectively capturing both aspects simultaneously. To address these limitations, we propose the Advanced Multimodal Sentiment Analysis with Dual Attention (A-MSDA) model, which integrates self-attention and cross-modal attention mechanisms in a unified dual-attention framework. This design enables robust multimodal fusion by extracting salient textual and visual features while modeling their image–text interaction comprehensively. Experimental evaluation on MVSA-Single and MVSA-Multiple datasets demonstrates that A-MSDA achieves notable improvements in accuracy and F1-score, outperforming existing techniques by up to 3.4% in F1-score on MVSA-Multiple, while maintaining competitive performance on MVSA-Single. These results highlight the potential of A-MSDA to advance research in deep multimodality and sentiment analysis.

Author 1: Soukaina FATIMI

Author 2: WAFAE SABBAR

Author 3: Abdelkrim BEKKHOUCHA

Keywords: Dual attention; self-attention; cross-modal attention; multimodal sentiment analysis; multimodal fusion; MVSA dataset; image–text interaction; deep multimodality

Paper 22: NDN-Based ICN Architecture Design to Improve Data Communication QoS in Kertajati Aerocity

Abstract: The development plan for Kertajati Aerocity requires adequate support from the data communication infrastructure. The primary challenge lies in the dense data flows, which involve very large volumes and diverse data types. The commonly used host-centric architecture suffers from latency, bandwidth utilization, and scalability issues, making it unsuitable for the dynamic Aerocity scenario. To address these limitations, this study proposes an Information-Centric Networking (ICN) architecture based on Named Data Networking (NDN), where communication relies on content names rather than IP addresses. The ndnSIM simulator was employed in the experimental evaluation of this architectural model, using operational data requirements derived from the Taoyuan Aerotropolis case. Performance was assessed through Quality of Service (QoS) metrics, including throughput, latency, and cache hit ratio. The simulation results indicate that throughput stabilized at ~10.1 Mbps (from 7.2 Mbps initially) with balanced node distribution, while latency averaged ~3 ms, with p95 < 10 ms and over 95% of requests completed within low-delay bounds despite initial spikes. Cache statistics were unavailable (CacheHits/Misses = 0) due to tracer settings and traffic patterns, so cache analysis is left for future work. These findings highlight the novelty of integrating NDN-based ICN with Aerocity-specific traffic requirements. The proposed model is presented as a scalable solution for the evolving data communication infrastructure of Kertajati Aerocity in the future, emphasizing designing an NDN-based architecture specifically designed for the multi-zone Aerocity ecosystem, including hierarchical and cross-zone naming schemes that model the operational flow of Aerocity.

Author 1: Enang Rusnandi

Author 2: Tri Kuntoro Priambodo

Author 3: Mardhani Riasetiawan

Keywords: Aerocity; Information-Centric Networking (ICN); Named Data Networking (NDN); Quality of Service (QoS); ndnSIM

Paper 23: Improving the Performance of TFS with Ensemble Learning for Cross-Project Software Defect Prediction

Abstract: Software defect prediction (SDP) plays a key role in improving software quality by identifying defect-prone modules early in the development cycle. While within-project prediction has been widely studied, cross-project defect prediction (CPDP) remains challenging due to differences in datasets, high feature dimensionality, and poor model generalization. To address these challenges, this study enhances the Transformation and Feature Selection (TFS) approach by integrating ensemble learning techniques. Three methods, Gradient Boosting Machine (GBM), stacking, and hybridization, were explored to evaluate their effectiveness in improving CPDP performance. Experiments were conducted using the AEEEM datasets, with preprocessing steps including normalization, feature reduction, and the Synthetic Minority Oversampling Technique (SMOTE) to handle data imbalance. The models were trained on source projects and tested on separate target projects, with the F1 score used as the main evaluation metric. Results show that the TFS × Stacking model achieved the highest overall performance, with a mean F1 score of 0.963, outperforming both TFS × GBM (0.958) and TFS × Hybridization (0.920). Compared to the original TFS × Random Forest method, the stacking approach consistently provided significant improvements across all project pairs. These findings highlight the potential of combining TFS with ensemble learning to enhance defect prediction in projects with limited or no historical data. This work not only advances CPDP research but also offers practical value to software teams by enabling more accurate identification of defect-prone modules and better allocation of testing resources.

Author 1: Pathiah Abdul Samat

Author 2: Yahaya Zakariyau Bala

Author 3: Nur Hamizah Hamidi

Keywords: Software; defect prediction; cross-project; ensemble learning; feature selection

Paper 24: From User Experience Evaluation to Design Guidelines for E-Commerce Websites

Abstract: The evolution of the Internet and advances in information technology have significantly transformed business practices, leading to widespread e-commerce website development. To succeed in a competitive environment, website owners and designers must ensure that their e-commerce websites provide positive user experiences (UX). This research proposes UX guidelines for e-commerce websites, comprising 27 categories: 20 pragmatic features and 7 hedonic features. The guidelines were developed by evaluating the UX of two local e-commerce websites (Shopee and Lazada) and two international websites (Amazon and Alibaba), using a questionnaire completed by 200 participants to collect both quantitative and qualitative data. Results revealed common UX problems across all websites, including hedonic issues such as unattractive design and pragmatic issues such as poor usability, unclear layout, lack of user-friendliness, non-intuitive design, poor performance, and disorganization. Unique UX problems were also observed: some hedonic and pragmatic issues were specific to local websites, while certain pragmatic issues were unique to international websites. For Amazon, quantitative analysis showed the highest average rating for hedonic and pragmatic metrics at 3.87 out of 5, while qualitative analysis identified the least number of UX issues (13 in total: 1 hedonic and 12 pragmatic).

Author 1: Layla Hasan

Author 2: Beatrice Lim Pei Ying

Keywords: User experience; UX guidelines; e-commerce websites; Malaysian e-commerce; hedonic features; pragmatic features

Paper 25: From External Stakeholder Pressure to Sustainable Practice in HEIs: Mechanisms and Internal Mediating Factors

Abstract: Sustainable procurement is an important part of sustainable development in HEIs, playing a pivotal role in optimizing resource allocation, fulfilling social responsibilities, and promoting green development. However, existing research has paid insufficient attention to the impact of external stakeholder pressure on HEIs’ sustainable procurement and its intrinsic action mechanism. To address this gap, this study aims to explore the influence path of external stakeholder pressure on HEIs’ sustainable procurement and identify key mediating factors. This study collected 260 valid data points from Chinese higher education institutions with more than one year of purchasing experience through snowball sampling. PLS-SEM analysis results show that external stakeholder pressure not only directly promotes sustainable procurement but also exerts an indirect effect through two mediating paths: affective commitment and professional knowledge. The mediating role of affective commitment is stronger than that of knowledge, and affective commitment itself has the strongest direct impact on sustainable procurement among all variables. Theoretically, this study enriches the application scenarios of stakeholder theory and institutional theory in the field of higher education sustainable management. Practically, it provides actionable references for HEIs to enhance sustainable procurement performance by strengthening external stakeholder collaboration, optimizing knowledge management systems, and fostering employees’ affective commitment.

Author 1: Xue Jin

Author 2: S. M. Ferdous Azam

Author 3: Jacquline Tham

Keywords: Sustainable procurement; HEIs; external stakeholder pressure; affective commitment; knowledge

Paper 26: Lexicon-Based Sentiment Analysis of Social Media Reviews for Floating Market Popularity in South Kalimantan

Abstract: Floating markets in South Kalimantan are culturally significant heritage destinations whose contemporary reputation is increasingly shaped by user-generated content on digital platforms. This study analyzes public perceptions of these markets by applying a lexicon-based sentiment analysis framework to 300 reviews collected from TripAdvisor and Google Maps between 2023 and 2024. The analytical workflow included text normalization, tokenization, stop-word removal, and stemming, with feature representation generated through term frequency–inverse document frequency (TF-IDF). Sentiment polarity was determined using bilingual lexicon-based scoring and categorized into positive, neutral, or negative sentiments. The results indicate that 45% of reviews expressed positive sentiment, highlighting cultural distinctiveness and riverfront experiences; 35% were neutral and provided descriptive logistical information; and 20% were negative, emphasizing waste issues, overcrowding, pricing inconsistencies, and perceived reductions in authenticity. TripAdvisor reviews exhibited greater emotional polarization than those on Google Maps. The findings demonstrate that lexicon-based sentiment analysis offers a transparent and effective approach for multilingual tourism contexts, providing insights into how digital narratives contribute to destination image formation. The study offers practical implications for improving environmental management, regulating visitor flows, and enhancing communication transparency within heritage tourism settings. It also contributes theoretically by underscoring the informational role of neutral reviews within electronic word-of-mouth dynamics. Future work may integrate machine learning-based sentiment classifiers or multimodal data to enhance analytical precision and extend the applicability of sentiment analysis in digital tourism research.

Author 1: Evi Lestari Pratiwi

Author 2: Ramadhani Noor Pratama

Author 3: Inayatul Ulya Ahyati

Author 4: Paula Dewanti

Keywords: Sentiment analysis; lexicon-based methods; floating markets; social media analytics; cultural tourism

Paper 27: AI-Driven Deep Learning Architectures for Robust Emotion Recognition

Abstract: Due to an insufficient labeled dataset, class-level variation emotion recognition becomes a challenging task in computer vision. Deep learning (DL) makes it possible to automatically learn meaningful patterns from facial expressions. It captures simple details such as edges, textures at low layers, and gradually builds up to more complex information, including Facial components and the overall meaning of the expression. Despite progress made via end-to-end learning, partial occlusions, inconsistent lighting, and biases within datasets are a few challenges that still remain. In this work, a DL based model is presented to classify two emotional states of human expression. The pipeline depends on several components, including the preparation of data, preprocessing and analysis, and the use of pretrained networks, dimensionality-reduction techniques, and region-based explanation via Grad-CAM. More than 2,000 images of happy and sad faces were derived from Kaggle. These images were used to test a custom-designed CNN and two widely adopted architectures, such as VGG16 and MobileNetV. The custom model attained an accuracy rate of 66% and 67% F1, while the VGG16 performed notably better with 78% accuracy and 77% F1, and the MobileNetV architecture, which achieved 77% accuracy and 73% F1. The statistical comparisons using paired t-tests and Wilcoxon signed-rank tests further confirmed these findings, showing that pre-trained models outperformed a custom CNN with a meaningful effect size. Although deeper networks are more susceptible to overfitting and the hand-crafted CNN suffered exhibited underfitting, the results indicate that pretained architecture provides a clear advantage for facial emotion recognition. This study makes a major contribution to existing computer vision research in removing the trade-off between accuracy and generalization, and opens doors to the application of lightweight yet interpretable models in practical affective computing systems.

Author 1: Hamad Ali Abosaq

Keywords: Deep learning models; computer vision; emotion recognition; image processing; Grad-CAM

Paper 28: User Experience Deficiencies in Mobile Tourism Applications: A Preliminary Study from Generation Z and Tourism Practitioners

Abstract: While previous research has broadly explored the user experience of mobile tourism applications (MTAs), few have examined user experience deficiencies from the dual perspectives of both tourism practitioners and Generation Z adults. This preliminary study sought to investigate user experience deficiencies present in MTA, with an emphasis on Generation Z adults. This study analysed the viewpoints of tourism practitioners and Generation Z users on MTA and revealed commonalities in their perspectives. Data was collected through semi-structured interviews with five tourism practitioners and eight Generation Z adults. Thematic analysis was used to identify key themes. The results revealed that the two groups of respondents held complementary perspectives on user experience deficiencies faced by MTA users. The findings reveal that numerous MTAs are plagued by perplexing information architecture, unappealing interfaces, and inadequate emotional resonance. This study transcends the systemic limitations and usability challenges identified in previous studies, concentrating on functionality issues, and offers initial practical suggestions for developers and designers to create user-centric interface designs.

Author 1: Huang Huihui

Author 2: Azliza Othman

Author 3: Nadia Diyana Mohd Muhaiyuddin

Keywords: Mobile tourism applications; Generation Z; tourism practitioners; user experience; preliminary study

Paper 29: Data-Driven Model for Optimizing Active Learning

Abstract: Data-driven models depend on extensive datasets for precise predictions; yet, acquiring adequate labeled data for training these models is a challenge, especially with medical datasets that are constrained by privacy considerations, resulting in a deficiency of labeled data. Active Learning (AL) has developed as a cost-effective strategy that minimizes the quantity of labeled data required for training by selecting the most informative samples. The performance of active learning methods is significantly influenced by data quality characteristics, and due to a lack of direction in selecting the most suitable active learning approach. The study presents a data-driven selection approach that suggests appropriate active learning methods based on dataset characteristics. The study examines the characteristics of the dataset and their impact on active learning performance, revealing significant correlations between data quality issues and the efficacy of active learning approaches. A rule-based selection model is subsequently constructed and verified by experiments and case studies across various datasets. The findings demonstrated consistent alignment between suggested and practically effective techniques. Statistical analysis verifies that the data-driven selection model exhibits reliability exceeding chance agreement, indicating its robustness and practical application in recommending AL techniques selection.

Author 1: Abang Asyraaf Aiman Abang Azahari

Author 2: Marshima Mohd Rosli

Author 3: Nor Shahida Mohamad Yusop

Keywords: Data-driven models; selection model; data quality characteristics; active learning

Paper 30: A Novel YOLO-Like Multi-Branch Architecture for Accurate Apple Detection and Segmentation Under Orchard Constraints

Abstract: This study introduces a novel YOLO-like multi-branch deep learning architecture designed for accurate apple detection and segmentation in orchard environments, addressing the persistent challenges of occlusion, illumination variability, and fruit clustering. The proposed model integrates an enhanced backbone with C2f modules and a Spatial Pyramid Pooling Fast (SPPF) block to capture multi-scale receptive fields, while a Feature Pyramid Network (FPN) combined with a Path Aggregation Network (PAN) ensures effective top-down and bottom-up feature fusion. To extend beyond bounding box localization, a prototype-based segmentation head is incorporated, enabling precise instance mask generation with reduced computational overhead. The model was comprehensively evaluated on the MinneApple dataset, consisting of high-resolution orchard images with polygonal annotations, and compared against state-of-the-art detection and segmentation frameworks, including Faster R-CNN, Mask R-CNN, SSD, YOLO variants, YOLACT, and SOLOv2. Quantitative results demonstrated that the proposed approach achieved superior mean Average Precision (mAP@0.5 = 0.76), precision (0.83), and F1-score (0.76), while maintaining a competitive inference speed of 40 FPS, confirming its suitability for real-time agricultural applications. Qualitative analysis further highlighted robustness in complex orchard conditions, reinforcing the model’s applicability for automated harvesting, yield estimation, and orchard monitoring. These findings advance the state of agricultural computer vision by unifying detection and segmentation in a lightweight, high-performance framework.

Author 1: Olzhas Olzhayev

Author 2: Nurbibi Imanbayeva

Author 3: Satmyrza Mamikov

Author 4: Bibigul Baibek

Keywords: Precision agriculture; detection; segmentation; YOLO-like architecture; multi-branch network; feature pyramid network; real-time inference; orchard monitoring

Paper 31: Kinematic Modeling and Design of a Robotic Manipulator System for Automated Fruit Harvesting in Intensive Orchards

Abstract: The development of robotic systems for automated fruit harvesting in intensive orchards has emerged as a critical response to labor shortages, high production costs, and the need for efficiency in modern agriculture. This study presents the kinematic modeling and design of a robotic manipulator system integrated into a mobile platform with an articulated lift mechanism, dual manipulators, and compliant gripping devices equipped with vision-based perception. The proposed system was modeled and validated through simulation in SolidWorks, enabling analysis of workspace coverage, kinematic stability, and motion optimization. Results indicate that the dual-manipulator configuration achieved a harvesting rate of up to 12 trees per hour, reducing the average fruit cycle time to less than seven seconds while lowering fruit loss to 14.5%, compared to over 30% in manual harvesting. The gripping device demonstrated a success rate of 94% with safe detachment forces between 2.5 and 3.5 N, ensuring minimal fruit damage and consistent quality. The lift mechanism provided stable vertical translation with minimal lateral deflection, supporting precise manipulator operation. Overall, the study highlights the potential of robotic manipulators to enhance productivity, safety, and sustainability in orchard management, while outlining future directions for field implementation, adaptive vision algorithms, and autonomous navigation.

Author 1: Nurbibi Imanbayeva

Author 2: Bekzat Amanov

Author 3: Aiman Nuramaganbetova

Author 4: Arman Moldashev

Author 5: Akbayan Aliyeva

Author 6: Sulukul Dauletbekova

Keywords: Robotic manipulator; kinematic modeling; fruit harvesting; intensive orchards; machine vision; deep learning

Paper 32: A Configurable Storytelling System for Emotion Recognition in Children: Design and Pilot Evaluation

Abstract: Computer games are very popular among children, with serious games being widely used for teaching specific skills and rehabilitation across various age groups. This work presents a configurable, story-based serious game designed to teach emotion recognition to young children (ages 6 to 11) through a peer-assisted learning approach. The system emphasizes pedagogical adaptability, allowing educators to customize content without programming knowledge. Its core innovation lies in structuring collaboration where older children co-create emotion-focused questions and facts for younger peers, operationalizing Vygotsky's social development theory. A pilot study (N=4) demonstrated the system's effectiveness in improving emotion identification, with 75% of participants showing increased engagement with the story's emotional cues. Results indicated successful peer-assisted interaction, where younger children answered most questions posed by older peers and showed critical engagement through fact verification. The findings suggest that this configurable, peer-assisted storytelling approach offers a promising, accessible method for fostering social-emotional learning in educational settings.

Author 1: Chadi Fouad Riman

Author 2: Wael Hosny Fouad Aly

Author 3: Carmen Mariana Pasca

Keywords: Computer gaming; serious games; education; emotions learning; peer-assisted learning

Paper 33: Leveraging AI and ML for Enhanced Business Intelligence Systems: A Research Landscape of Trends, Influences, and Future Directions

Abstract: This research provides a systematic review of AI and ML applied to the BI context from 2014 to 2024. By characterizing the article and citation distribution and by tracing the topics of publications over time, this study outlines major developments, landmark contributions, and the future outlook of the research discipline at hand. However, despite rapid advancements, existing research remains fragmented. It lacks a consolidated understanding of how AI and ML are shaping modern BI systems, creating an urgent need for an integrated review. Prior studies often examine isolated techniques or industry-specific implementations, but very few provide a comprehensive synthesis that maps long-term trends, methodological patterns, and unresolved challenges in AI/ML-enabled BI. The study reviews more than 200 research articles obtained from top academic databases and finds a rise in the integration of AI/ML with conventional BI systems. Our results show the emergence of real-time, predictive, and automotive analytics as much-needed and valuable to consumers. Artificial Intelligence strengthens the capabilities for visualizing data, so complex information becomes easier to access by those who need it. Research literature shows an increase in publications about AI and ML applications in BI systems because these technologies have gained substantial practical importance. The field of study investigates the ways AI and ML improve BI systems by paying special attention to predictive analytics as well as decision-making processes. The main aspects of interest unite advanced AI implementations with user-centric tools that serve multiple industries. Future directions for researchers are AI ethics in BI, as well as creating simple AI tools for non-programmers and investigating the influence of AI-based BI systems on different industries in the long run. This study also presents future research suggestions.

Author 1: Sunil Mandaliya

Author 2: Priti Kulkarni

Keywords: Artificial intelligence; machine learning; business intelligence; digital transformation; data analytics; bibliometric analysis

Paper 34: Performance Evaluation of Software Multi-Threshold Decoders for Self-Orthogonal Codes in Modern Broadband Wireless Communication Systems

Abstract: 5G and Internet of Things (IoT) wireless systems face challenges to reliable data transmission due to multipath fading, intersymbol interference, and the need for low-complexity Forward Error Correction (FEC). Conventional FEC techniques, such as Low-Density Parity-Check (LDPC) and turbo codes, provide high reliability but are unsuitable for resource-constrained IoT devices due to high decoding complexity. The aim of this study is to evaluate Multi-Threshold Decoders (MTDs) applied to Self-Orthogonal Codes (SOCs) as a low-complexity FEC solution in Orthogonal Frequency-Division Multiplexing (OFDM) and Multiple Input Multiple Output (MIMO) systems with Space–Time Coding (STC). The systems are modeled under ITU-R (Outdoor A, TU6, RA6) and 3GPP Spatial Channel Model (Urban Macro/Micro) fading environments and compared with LDPC (WiMAX, DVB-S2) and turbo codes in terms of Bit Error Rate (BER), Signal-to-Noise Ratio (SNR), decoder complexity, antenna diversity, modulation order, and throughput. Results indicate that SOC+MTD outperform short LDPC and turbo codes under deep fading while achieving reliability comparable to long LDPC codes at significantly lower decoding complexity. Min-sum refinement and approximate Maximum-Likelihood (ML) detection provide up to 2 dB additional SNR gain, and 1×3 antenna diversity reduces required Eb/N₀ by ~7 dB at BER = 10⁻⁵. Higher-order modulations such as 8PSK and 16APSK achieve 1.5–2× higher bit rates with moderate SNR penalties, while Open Computing Language (OpenCL) based Graphics Processing Unit (GPU) acceleration enables a 32-fold increase in simulation speed. These findings demonstrate that SOCs decoded with MTD represent a promising low-complexity, high-reliability FEC approach for 5G and IoT physical layers.

Author 1: Nurlan Tashatov

Author 2: Gennady Ovechkin

Author 3: Zhuldyz Sailaukyzy

Author 4: Eldor Egamberdiyev

Author 5: Dina Satybaldina

Author 6: Gulmira Danenova

Author 7: Zarina Khassenova

Keywords: Forward Error Correction; Multi-Threshold Decoder; Self-Orthogonal Codes; Orthogonal Frequency-Division Multiplexing; Multiple Input Multiple Output; Bit Error Rate; signal-to-noise ratio; LDPC; turbo codes; space–time coding; fading channels; GPU acceleration; OpenCL

Paper 35: CICA Framework: Harnessing CSR, AI, and Blockchain for Sustainable Digital Culture

Abstract: Digital transformation has created new opportunities for organizations, but it has also intensified cybersecurity risk. In emerging economies, where institutional support and digital literacy remain limited, cybersecurity awareness plays a crucial role in strengthening digital resilience and fostering a sustainable digital culture. This study introduces the CSR-Integrated Cybersecurity Awareness (CICA) Framework, which conceptualizes Corporate Social Responsibility (CSR) as a key driver of cybersecurity awareness, reinforced by the adoption of artificial intelligence (AI) and blockchain technologies. Data were collected from companies in Central Java, Indonesia, that implement CSR-based digital initiatives, with responses gathered from managers, CSR officers, and IT staff. Using Structural Equation Modeling (SEM), the findings show that CSR significantly enhances cybersecurity awareness, AI adoption strengthens proactive security measures, and blockchain increases trust and transparency. The results also reveal that CSR mediates the relationship between digital technology adoption and sustainable digital culture. This study contributes by integrating CSR and cybersecurity through emerging technologies, offering theoretical insights and practical implications for organizations in developing regions.

Author 1: Danang Danang

Author 2: Agustinus Budi Santoso

Author 3: Maya Utami Dewi

Keywords: Cybersecurity awareness; corporate social responsibility; artificial intelligence; blockchain; sustainable digital culture; emerging economies

Paper 36: Recursive Gated Convolution-Based YOLOv11 Framework for Operator Safety Management in Live-Line Work

Abstract: In live-line work scenarios, it is essential for workers to wear electric field shielding clothing to prevent fatal accidents caused by electric shock. Accordingly, this study developed an electric field shielding clothing detection system for live-line working environments based on the YOLOv11 framework. Previous research has explored intelligent wearable detection systems for personal protective equipment such as safety helmets. However, compared to safety helmets, electric field shielding clothing comes in more varieties and is more challenging to identify. To address the challenges mentioned above, this study constructed a dual-layer detection model for operator detection and electric field shielding clothing detection in live-line work scenarios. The first layer employs an improved detection transformer (IDETR) to locate operators within the environment. The second layer, based on the YOLOv11 framework integrated with recursive gated convolution (GnConv), is designed to classify three types of personal protective equipment, including electric field shield clothing, electric field shield masks, and electric field shield gloves. Finally, the experimental results showed that compared with the DETR, the accuracy of the IDETR-based worker localization model improved by 2.29%. The accuracy of the GnConv-based YOLOv11 framework in the electric field shielding clothing detection task reaches 90.40%.

Author 1: Dapeng Ma

Author 2: Liang Yang

Author 3: Kang Chen

Author 4: Feng Yang

Author 5: Ao Cui

Author 6: Rundong Yang

Author 7: Zhilin Wen

Author 8: Donghua Zhao

Keywords: Detection transformer; recursive gated convolution; YOLOv11; personal protective equipment; live-line work scenarios

Paper 37: Evaluating the Effectiveness and Usability of Microsoft Threat Modelling Tool in Undergraduate Cybersecurity Education

Abstract: As cyber threats evolve, equipping students with hands-on experience in identifying and mitigating system vulnerabilities is critical for developing a cybersecurity-aware workforce. There are a variety of threat modelling tools available on the market, and it is challenging for educators to select the best tool for their students to learn and identify any possible threats that may exploit system vulnerabilities. This study investigates the effectiveness and usability of the Microsoft Threat Modelling Tool (MTMT) among undergraduate students, addressing the need for a practical tool in cybersecurity education. This study was conducted in four phases. The first phase involves conducting a comprehensive literature review to understand the features, capabilities, and tools of the threat modelling tools being compared, specifically the MTMT. Phase two consists of defining the evaluation criteria for assessing the tool's effectiveness and usability. Criteria for error frequency, ease of use, and user-friendliness will be developed, with particular focus on their relevance to educational environments, especially for undergraduate students. Phase three involved data collection, during which participants were recruited and had hands-on sessions with the tool. Training sessions were conducted using case studies to familiarise participants with the tool's features and functionalities. The last phase involves developing assessments to evaluate participants’ knowledge, effectiveness and usability of the tools. The evaluation includes structured usability testing and post-assessment of students’ knowledge and skill acquisition. Findings reveal that MTMT enhances students’ comprehension of threat modelling concepts, bridging the gap between theoretical knowledge and real-world cybersecurity practices. However, the study also highlights areas for improvement in the tool’s interface and documentation to better support student learning. These insights enhance educational strategies, foster active learning, and equip students for real-world cybersecurity challenges. The results emphasise the tool’s potential to strengthen the integration of threat modelling into the cybersecurity field, thereby fostering essential skills for safeguarding organisational and digital infrastructures. The novelty of this study lies in the methodology used to measure the effectiveness and usability of the threat modelling tool. The tool’s effectiveness was measured using the effectiveness formulas from ISO/IEC 25022:2016(E), while its usability was measured using the System Usability Scale (SUS).

Author 1: Nor Laily Hashim

Author 2: Ahmad Zuhairi Bin Mohd Yusri

Keywords: Cybersecurity education; threat modelling; stride; usability testing

Paper 38: Robust Detection of Partially Occluded Faces in Low-Light Scenarios Using YOLOv7 and YOLOv6

Abstract: Partial occlusion and low light are significant challenges for face detection, limiting its effectiveness in critical applications such as security, surveillance, and user identification within computer vision. This study evaluates the effectiveness of two influential deep-learning models, YOLOv6 and YOLOv7, in identifying partially occluded faces in uncontrollable, real-world conditions. Training both models and assessing them with the help of comprehensive data-augmentation schemes that facilitate the occurrence of generalization, a carefully selected sample of partially blocked and hidden images of faces was used in all the experiments performed under low-light exposure. Findings indicate that YOLOv7 systematically outsmarts YOLOv6 in all key measures of performance, including precision (0.92 vs. 0.90), recall (0.89 vs. 0.79), as well as the mean Average Precision (mAP), which proves its ability to recognize hidden faces under adverse environments better. YOLOv7 takes a longer time to be trained, but with its enhanced design, especially the Extended Efficient Layer Aggregation Network (E-ELAN), feature extraction and real-time detection become much smoother. The statistics of this research indicate a visible increase, which indicates that YOLOv7 is reasonably suitable to be implemented in real-life, where a strong ability of face recognition, even with occlusions and low visibility, is required. This study contributes to the advancement of face detection technologies, tackling developing privacy and security needs in increasingly masked and low-visibility spaces.

Author 1: Nayef Alqahtani

Author 2: Amina Shaikh

Author 3: Imran Khan Keerio

Keywords: Deep learning; partial occlusion; YOLOv6; YOLOv7; real-time detection; low-light face recognition

Paper 39: Semantic Segmentation Algorithm of Animal Husbandry Image Based on an Improved U⁃Net Network

Abstract: Due to the limitations of unclear edges and fuzzy features in image segmentation tasks, this study proposes an enhanced U⁃Net semantic segmentation network utilizing the local and global fusion attention module in response to the drawbacks of fuzzy features and unclear edges in image segmentation tasks. Firstly, a feature extraction module combining convolution and Transformer is introduced in the bottleneck layer, so that the network can fully simultaneously capture local and global features, and effectively promote the fusion of local and global features. Secondly, the CBAM attention module is added to the skip connections between the encoder and decoder. Finally, the output feature map is processed using the ASPP module to enhance focus on target features and improve segmentation performance. Experiments conducted on four animal husbandry segmentation datasets show that the LCA_Net model proposed in this study achieves an IoU score of 90.19% and a Dice score of 94.83%, compared with U-Net and other mainstream segmentation networks, it has improved. This study offers effective technical support for advancing aquaculture status monitoring and lays a foundation for further development in this field.

Author 1: Jia Li

Author 2: Jinjing Zhang

Author 3: Fengjiao Jiang

Keywords: Machine vision; semantic segmentation; feature fusion; attention mechanism

Paper 40: AI-Driven Professional Profile Categorization and Recommendation System

Abstract: The exponential growth of applications in digital and information system domains has made the identification of qualified candidates increasingly complex, resulting in longer and less efficient recruitment processes. Recruiters frequently deal with heterogeneous and unstructured résumés, which complicates skill assessment and increases the risk of mismatches between candidates and job requirements. To address these challenges, this research proposes an AI-based framework for the automatic classification and recommendation of professional profiles using natural language processing (NLP), text mining, and supervised machine learning techniques. The methodology includes the comparative evaluation of several classification algorithms—Logistic Regression, Random Forests, Support Vector Machines (SVM), k-Nearest Neighbors (KNN), Gradient Boosting (GB), and Naïve Bayes—to identify the most accurate and robust model. The framework also incorporates a similarity-based matching mechanism to align candidate profiles with job postings. Experimental results show a classification accuracy of 96.38%, demonstrating the model’s effectiveness in enabling faster, more reliable, and objective recruitment decisions while providing candidates with insights into their compatibility with labor market expectations.

Author 1: Marouane CHIHAB

Author 2: Hicham BOUSSATTA

Author 3: Mohamed CHINY

Author 4: Nabil Mabrouk

Author 5: Younes CHIHAB

Author 6: Moulay Youssef HADI

Keywords: Professional profile classification; profile recommendation; natural language processing (NLP); supervised learning; Logistic Regression; Random Forest; Support Vector Machine (SVM); k-Nearest Neighbors (KNN); Gradient Boosting; Naïve Bayes; AI-based recruitment

Paper 41: Robotic Process Automation (RPA) Scripting Model Using Machine Learning (ML) for Enterprise Data Validation and Integration

Abstract: This research presents an automated data processing model based on RPA Scripting, designed to enhance efficiency in extracting, validating, and integrating information from various web platforms. The automated workflow begins with the use of a tool that simulates human interaction on web platforms to obtain data automatically and reliably. The data is then organized and cleaned using processing techniques that prepare it for analysis. As a key component of the model, Machine Learning algorithms have been incorporated to detect errors, identify unusual patterns, and classify records, thereby improving data quality before storage. Finally, the processed data is loaded into a database and visualized through a dynamic dashboard that supports decision-making via reports and indicators. In conclusion, integrating Machine Learning algorithms within an RPA Scripting model not only optimizes the execution of automated tasks but also equips the model with intelligence to anticipate errors and adapt to changes in the data. This enables the development of a more robust, reliable, and adaptive automated process, aligned with current requirements for real-time analysis and decision-making.

Author 1: Luis Ángel Bendezú Jiménez

Author 2: Jorge Luis Juan de Dios Apaza

Author 3: Ruben Oscar Cerda García

Keywords: RPA; scripting; data automation; Machine Learning; data validation; data integration; intelligent workflows

Paper 42: How did the Intelligent Search Engine Become Popular Among Chinese Residents Since the Emergence of Deepseek?

Abstract: This study, based on the Technology Acceptance Model (TAM) and the Theory of Planned Behavior (TPB), combined with the Stimulus-Organism-Response (SOR) theoretical framework, explores the influencing factors and promotion mechanisms for the popularization of “Deepseek” intelligent search engines in China. Through a questionnaire survey, 343 valid samples were collected, and the structural equation model (SEM) was used for data analysis, verifying 8 out of 10 research hypotheses. Perceived usefulness (PU) and perceived ease of use (PEOU) significantly and positively influence the intention to use, indicating that users' cognition of the functional value and operational convenience of AI search engines is a key driving factor. Subjective norms (SN) directly promote the intention to use but do not indirectly affect it through perceived usefulness, suggesting that the influence of the social environment stems more from irrational paths (such as group atmosphere infection). Perceived behavioral control indirectly enhances the intention to use by improving perceived ease of use and usefulness, highlighting the importance of users' own abilities and device support. Perceived playfulness only indirectly affects the intention to use through perceived usefulness and has no significant effect on ease of use, possibly because the entertainment function increases operational complexity. Technological facilitation has a significant positive impact on both perceived usefulness and ease of use, indicating that the optimization of technical performance (such as interaction efficiency and powerful functions) is the core for enhancing user experience.

Author 1: Zhixuan Wang

Author 2: Yuying Song

Author 3: Hanyu Guo

Keywords: AI search engine; Deepseek; SOR; Technology Acceptance Model; Theory of Planned Behavior

Paper 43: Software Project Effort Estimation Using Formal Method and Model Checker

Abstract: Software project effort estimation is a critical component of software development, as it determines the time and financial resources required to complete a project. Existing estimation techniques—ranging from empirical models and algorithmic methods to heuristic and expert-based approaches—struggle with inconsistent accuracy due to the inherent complexity, subjectivity, and contextual variability across software projects. Although earlier formal methods aimed to reduce confusion by being very precise, they usually don't allow for automated logical analysis or check if the assumptions used for estimation are consistent with one another. To address these limitations, this study introduces a novel formal modeling framework that integrates Z-Specification with the Z3 SMT solver to both formalize and computationally verify effort estimation models. The use of Z notation guarantees the meaning is precise and unambiguous. Furthermore, SMT (Satisfiability Modulo Theories) reasoning adds powerful new abilities that older methods lacked. These new capabilities include automatically finding constraint violations, confirming how parameters depend on one another, and determining feasible estimation ranges under clearly defined conditions. This integration not only reduces ambiguity but also provides a verifiable, machine-checkable basis for evaluating, refining, and comparing diverse effort estimation methods, thereby offering a more robust foundation than traditional or solely formalized models.

Author 1: Abdulaziz Alhumam

Keywords: Software effort estimation; cost estimation; formal methods; SMT Solver; automated verification; Z-specifications

Paper 44: Cross-Lingual Sentiment Analysis in Low-Resource Languages: A Recent Review on Tasks, Methods and Challenges

Abstract: Cross-lingual sentiment analysis (CLSA) has become increasingly important in natural language processing and machine learning, enabling the understanding of opinions across diverse linguistic communities, particularly in low-resource languages (LRLs). Despite growing attention, persistent challenges such as limited annotated data, semantic misalignment, and cultural variation in sentiment expression continue to hinder progress. This systematic literature review (SLR) examines recent developments by analyzing the tasks, methods, and challenges reported in CLSA studies focused on LRLs. Following the PRISMA 2020 framework, a comprehensive search was conducted across major databases, including Scopus, IEEE Xplore, SpringerLink, Elsevier, and Google Scholar, covering studies published between 2021 and 2025. After applying inclusion and exclusion criteria, 27 studies were selected for analysis. The findings reveal that while polarity detection remains the dominant sentiment analysis task, emerging directions such as aspect-based sentiment analysis (ABSA), emotion detection, and hate speech recognition are gaining traction. Methodologically, most studies rely on multilingual pre-trained language models (PLMs), supplemented by machine translation, transfer learning, few-shot learning, and hybrid approaches. However, key challenges remain, including the scarcity of high-quality datasets, instability of few-shot performance, difficulties in handling dialectal variation, bias in PLMs, and the lack of standardized evaluation benchmarks. This review concludes by emphasizing the need for more culturally grounded tasks, adaptive hybrid frameworks, and fairness-aware evaluation practices to build robust cross-lingual frameworks and richer linguistic resources for underrepresented languages.

Author 1: Nor Zakiah Lamin

Author 2: Azwa Abdul Aziz

Keywords: Cross-lingual sentiment analysis; low-resource language; natural language processing; pre-trained language models; transfer learning; few-shot learning

Paper 45: Advances in Natural Language Processing for Radiology: State-of-the-Art Techniques, Applications, and Open Challenges

Abstract: Radiology reports encode critical clinical observations from medical imaging in an unstructured textual form that is central to modern clinical diagnosis and decision support. In this context, natural language processing (NLP) has emerged as a key clinical NLP technology for automatically extracting, classifying, and interpreting information from radiology reports. This study presents a structured review of more than sixty recent contributions on NLP for radiology, covering approaches that range from traditional rule-based pipelines to contemporary deep learning and transformer-based models. We examine how deep learning architectures, including BERT, GPT-4, multimodal transformers, and vision–language alignment networks, are applied to core tasks such as disease classification, tumor response assessment, cancer phenotype extraction, radiology report generation, cohort identification, quality assurance, and longitudinal patient follow-up. Particular attention is given to knowledge graph integration, multimodal cross-attention, and zero-shot learning strategies that adapt large language models to radiology-specific workflows. We also analyze key barriers to clinical adoption, including limited annotated data, domain generalization gaps across institutions, ethical and fairness concerns, and the need for transparent model explainability. Based on this synthesis, the review outlines future research directions for building interpretable, multimodal, and clinically robust NLP solutions that integrate technological, clinical, and operational perspectives to advance radiology report analysis and medical imaging–driven care.

Author 1: Kotha Chandrakala

Author 2: Shahin Fatima

Keywords: Natural language processing; radiology reports; deep learning; transformers; medical imaging; report generation; clinical NLP

Paper 46: AI-Driven Multimodal Frameworks for Cardiovascular Diagnostics: Integrating Sensors, Imaging, and Robotic Systems

Abstract: Cardiovascular disease is still the leading cause of death, and a definitive cure has not yet been found, so this is the time to make important changes in prevention and early diagnosis. Integrating artificial intelligence, machine learning, wearable sensors, and biomedical imaging is changing healthcare for cardiovascular care. Recent breakthroughs have reported that the use of smart immune sensors and artificial intelligence helps diagnose disease by testing blood and urine samples. The upcoming advances from various emerging technologies are expected to greatly enhance the overall accuracy and personalization of diagnostic processes within the medical field. This essay explores difficulties relating to domain adaptation, variability in data, and interpretability, including the need for rigorous validation tests and ethical considerations. The new system is made up of several programs that help the user to make decisions more efficiently in situations where rapid action is needed, while considering privacy preservation, clinical quality improvement, and energy efficiency. This review of more than sixty recent studies is an attempt to broaden the field of cardiovascular care by introducing a roadmap for further research. The creation of a fully responsive cardiovascular diagnostic system is not yet complete and requires the contribution of several entirely different fields of science.

Author 1: Chandrasekhara Reddy T

Author 2: Ramesh Babu P

Keywords: Cardiovascular disease (CVD); machine learning (ML); artificial intelligence (AI); wearable sensors; deep learning; Biomedical Signal Processing; microfluidics; nano sensors; robotic intervention; medical imaging; predictive modelling; causal inference; domain adaptation; federated learning; clinical validation

Paper 47: Sustainable IoT Smart Home Perceptions Across Demographic and Vulnerable User Segments

Abstract: The rapid expansion of Internet of Things (IoT) technologies has transformed smart home systems into essential tools for enhancing safety, independence, and quality of life, particularly for older adults and individuals with disabilities. However, the extent to which these groups understand and adopt IoT-enabled smart homes remains limited. This study addresses insufficient knowledge and uneven adoption of smart home IoT technologies among vulnerable demographic groups, examining how demographic factors shape levels of awareness and readiness for use. Two parallel approaches were employed for a validated questionnaire (of 15 sections) distributed to 249 participants and in parallel with semi-structured interviews for 25 selected individuals in Najran, Saudi Arabia, during the summer of 2023. Quantitative data were gathered and analyzed using descriptive statistics and multivariate analyses of variance. Qualitative data were under content analysis. Results revealed low levels of knowledge regarding IoT-enabled smart home systems among the respondents and different groups. Significant differences were found for groups of different employment status, age (15 to 30 and 30 to 45 years), economic status (above average), and disability status. However, no significant differences were found for gender or marital status. Qualitative insights indicated major concerns related to affordability, user passivity, lack of technical support, poor internet connectivity, device overload, issues of privacy, security, and system reliability. It was highly recommended to perform targeted measures to improve awareness, accessibility, inclusive design, and infrastructure. Future work may address spatial and temporal variations in IoT adoption, develop tailored training models for older adults and individuals with disabilities, and assess the effectiveness of policy initiatives aimed at increasing smart home readiness. These efforts can further improve the safe and effective integration of IoT technologies and improve indoor life quality and sustainability for vulnerable people.

Author 1: Burhan Mahmoud Hamadneh

Author 2: Zeyad Alshboul

Author 3: Nabhan Mahmoud Hamadneh

Author 4: Turki Mahdi Alqarni

Author 5: Malek Turki Jdaitawi

Keywords: Internet of Things; smart home; vulnerable people; smart home; energy efficiency

Paper 48: The Evolution of Hackathons as an Innovation Tool: A Systematic Analysis

Abstract: Hackathons have established themselves as dynamic open innovation spaces that promote interdisciplinary collaboration and creative problem-solving. This systematic review, which follows the PRISMA methodology, synthesizes the findings of 73 articles from Scopus, Web of Science, and IEEE Xplore, with the aim of analyzing the evolution, impacts, and knowledge gaps surrounding hackathons as an innovation tool. The study identifies a growing trend in their global implementation, with a particular emphasis on skill development, driving innovation, and strengthening entrepreneurial competencies. However, limitations are evident, such as the scarcity of longitudinal studies, the poor assessment of their long-term sustainability, and the geographical concentration of research in technologically advanced countries. Future research should focus on comparing organizational models, measuring long-term results, and including diverse contexts. Our findings underscore the potential of hackathons to boost creativity and entrepreneurship, as well as foster sustainable and collaborative innovation processes.

Author 1: Claudia Marrujo-Ingunza

Author 2: Meyluz Paico-Campos

Keywords: Hackathon; evolution; tool; innovation; review

Paper 49: SBERT-Based Stacking Ensemble Model for Fake News Detection

Abstract: Fake news has become a significant global challenge, affecting public opinion, social dynamics, and decision-making processes. Detecting fabricated news accurately and efficiently remains a challenging task due to the diversity of content, writing styles, and subtle semantic nuances. In this study, we propose a stacking ensemble model that uses SBERT-based semantic embeddings to improve the detection of fake news. The model integrates several machine-learning classifiers with a meta-learner to enhance robustness and predictive reliability. Experiments on the WELFake dataset show that the proposed model achieves 92.74% accuracy, a 93.01% F1-score, and a 97.93% ROC-AUC in classifying fake and real news. These results demonstrate the model’s effectiveness and suggest its potential for broader application across different languages and news domains.

Author 1: Abdulaziz A Alzubaidi

Author 2: Amin A Alawady

Keywords: Fake news detection; machine learning; SBERT embeddings; stacking ensemble; Random Forest; Logistic Regression; MLP; XGBoost

Paper 50: Aligning Confidence and Localization: An Enhanced DINO Model for Small Object Detection in Robotic Nursing

Abstract: In care environments such as nursing homes, robots performing tasks (e.g., feeding assistance) must accurately identify and locate target objects to ensure safe and efficient execution. However, real-world applications face challenges such as numerous small objects, complex backgrounds, and severe occlusions, all of which compromise detection performance. To address these challenges, this study proposes an end-to-end object detection algorithm based on an enhanced DINO framework. Transformer-based DINO is adopted as the baseline, leveraging its global modeling capabilities and avoiding the complex pre- and post-processing required by traditional CNN detectors. In addition, an improved Align-Loss is introduced to enhance small-object detection and address misalignment issues within DINO. Furthermore, a GhostConv module is integrated into DINO’s ResNet50 backbone to reduce the computational load of feature extraction and accelerate detection. Finally, multi-scale data augmentation and transfer learning are applied during training to improve detection accuracy and accelerate convergence. To validate the proposed method, experiments were conducted on the augmented MYNursingHome dataset and the COCO dataset. On the MYNursingHome dataset, the proposed approach improved mAP by 3.1% and APs by 2.6% over the DINO baseline, while reducing parameters from 47M to 39.6M and FLOPs from 279G to 243G. On the NVIDIA Jetson Orin Nano Super, inference speed increased from 16.8 FPS to 18.9 FPS (+12.5%). The experimental results demonstrate that the improved DINO detector proposed in this study exhibits a significant advantage in small object detection for nursing scenarios, providing algorithmic support for intelligent and efficient robotic care.

Author 1: Yanchen Du

Author 2: Qingzhuo Yuan

Author 3: Shengli Luo

Author 4: XiaoLong Shu

Author 5: Xu Wang

Author 6: Yuheng Jiang

Author 7: Hongliu Yu

Keywords: Small object detection; matching alignment; robotic object detection; nursing feeding; DINO; GhostConv; edge deployment; transformers

Paper 51: Sleep Quality and Burnout Syndrome in Students at a University in Lima, Peru: A Cross-Sectional Study

Abstract: The World Health Organization (WHO) reports that 15% of mental health problems develop in people with demanding work and academic conditions. Both sleep quality problems and Burnout Syndrome (BS) are recognized as significant problems in university settings. In Lima, Peru, the situation is critical, as BS affects up to 60% of university students. Therefore, quantifying this problem through a nursing intervention is crucial. The objective of this study was to determine the relationship between sleep quality and BS in university students at a university in Lima. Using a quantitative, cross-sectional, and correlational approach, the Pittsburgh Sleep Quality Index (PSQI) and the Student Burnout Scale (EUBE) were applied to a sample of 216 nursing and systems engineering students, using Spearman's Rho test and the Multinomial Logistic Regression Model. The findings revealed a moderate negative correlation between sleep quality and SB (Rho=-0.508; p < 0.001) and a relationship between sleep quality problems and mild SB (RRR=7.84565; p = 0.005). Furthermore, 85.19% of participants experienced sleep problems that warranted medical attention and treatment, and 86.57% had mild SB. Sleep quality problems and the development of SB are prevalent in this population; therefore, it is essential to continue studying them and integrating specific intervention strategies.

Author 1: Yajaira Garay-Castro

Author 2: Luis Acosta-Avila

Author 3: Ana Flores-Hiyo

Author 4: Ana Huamani-Huaracca

Author 5: Sebastián Ramos-Cosi

Author 6: Gina León-Untiveros

Author 7: Alicia Alva-Mantari

Keywords: Burnout Syndrome; sleep quality; university students; mental health

Paper 52: A Conceptual Model for an Ontology-Based Dietary Recommendation Plan in Crohn’s Disease

Abstract: Crohn's Disease (CD) is a long-term inflammatory bowel disorder that affects the digestive system. It is influenced by geography, diet, genetics, and immune response. Patients often experience difficulties managing CD due to the complexity and heterogeneity of the condition. Despite increasing scientific efforts, knowledge within the domain remains scattered and fragmented across different concepts. The purpose of this study is to develop a conceptual model of CD using knowledge engineering to organize domain knowledge and clarify the main aspects and their relationships. The conceptual model is created following the Ontology Development 101 methodology to define the creation of classes, properties, and restrictions. The proposed model is designed to organize and integrate multiple aspects of the condition, such as symptoms, treatments, risk factors, and patient profiles, with a particular emphasis on dietary recommendations. The resulting model consists of eight primary classes and fifteen key relationships, clarifying the connections between patients, symptoms, treatments, and diagnosis. The CD conceptual model does not comprehensively address the genetic, environmental, cultural practices, or neurobiological factors related to CD. To address these limitations, future work should focus on integrating real-world clinical data and considering broader demographic contexts. This study examines nutritional treatments for CD, such as Exclusive Enteral Nutrition (EEN), the low-FODMAP diet, and the Crohn’s Disease Exclusion Diet (CDED) that emphasize the role of diet in personalized healthcare. The CD conceptual model developed is the basis for building a comprehensive ontology-driven system that will help Crohn's patients with personalized dietary advice and future decision support systems that will improve their clinical care.

Author 1: Rezan Almehmadi

Author 2: Aisha Alsobhi

Author 3: Omaima Almatrafi

Keywords: Conceptual model; ontology; dietary recommendations; Crohn’s Disease; domain knowledge

Paper 53: Swarm Intelligence-Based Optimization of FACTS Devices: A Review of Operation, Control and Emerging Algorithms

Abstract: The integration of Flexible AC Transmission System (FACTS) devices into modern power networks plays a pivotal role in enhancing voltage stability, reducing transmission losses, and improving overall power transfer capability. Determining the optimal location and sizing of these devices is a critical task that significantly influences system performance. In recent years, swarm intelligence (SI) algorithms have emerged as powerful optimization tools for addressing such complex, nonlinear, and multi-objective problems in power systems. This study presents a comprehensive review of the application of swarm intelligence techniques, Artificial Bee Colony (ABC), Bacterial Foraging Optimization (BFO), Dragonfly Algorithm (DA), Salp Swarm Algorithm (SSA), and Particle Swarm Optimization (PSO). These algorithms are used to optimize the placement and sizing of FACTS devices, such as Static Var Compensators (SVCs), Thyristor-Controlled Series Capacitors (TCSCs), and Static Synchronous Compensators (STATCOMs). The review highlights the underlying mechanisms, strengths, and limitations by comparing the performance of each algorithm in terms of convergence, optimal location, and sizing of a particular FACT device in a power transfer system to enhance voltage stability, minimize real power losses, and improve system loadability. The review provides a comprehensive resource for researchers and practitioners interested in applying swarm intelligence-based optimization techniques of FACTS devices in power transmission systems.

Author 1: Patricia Khwambala

Author 2: Kumeshan Reddy

Author 3: Senthil Krishnamurthy

Keywords: Swarm intelligence; FACTS devices; power transfer system; voltage stability; power losses

Paper 54: SPA-DCN-NET: A Gated Multi-Scale Local Contrast Normalization Network for Ultrasound Image Segmentation of Liver

Abstract: Segmentation of the liver in ultrasound images is a critical task in medical image analysis, yet it remains challenging due to acoustic speckle noise, brightness instability, and deformations caused by probe pressure. To address these problems, this study presents SPA-DCN-NET, a lightweight framework that integrates three synergized components: first, a learnable gated local contrast normalization (Gated LCN) module utilizes a sigmoid soft-gate mechanism to dynamically fuse LCN-enhanced features with original features, effectively stabilizing the features for training. Second, a spatial pyramid attention (SPA) module applies multi-scale context aggregation, transforming the clear features provided by Gated LCN into deformable convolutional networks. Third, these features guide deformable convolutional networks to adaptively adjust sampling grids, ensuring precise delineation of irregular liver boundaries in ultrasound images. Experimental results demonstrate that our SPA-DCN-NET achieved mean IoU scores of 83.52%, 75.57%, 73.85%, and 85.94% across the four datasets, respectively. These results are all higher than those obtained by UNet, nnUNet, and ResUNet. The metrics indicate that our SPA-DCN-NET is more adaptable to the ultrasonic medical environment compared to other existing medical segmentation networks, and it is a recommended network of image analysis for ultrasound abdominal scans.

Author 1: Su Ming Jian

Author 2: Afizan Bin Azman

Author 3: Afizan.Azman@taylors.edu.my

Keywords: Spatial pyramid attention; deformable convolutional network; ultrasound image; medical image analysis; soft-gate mechanism; local contrast normalization

Paper 55: Generalizing In-Field Plant Disease Diagnosis: A Deep Transfer Learning Approach for Multi-Crop and Heterogeneous Imaging Conditions

Abstract: Plant diseases pose a serious threat to agricultural productivity, which can cause significant crop losses if not addressed quickly and appropriately. There are significant opportunities for digital image-based treatment with computer vision and artificial intelligence. The main challenges in recognizing image-based plant diseases are: developing a single model capable of diagnosing diseases in various types of plants. Ensuring the model remains reliable even when images are taken under varying lighting conditions, backgrounds, and camera quality. In addition, the challenge in this study is to present a model capable of recognizing leaf diseases of multiple food crops, especially rice and corn. The purpose of this study is to identify leaf diseases of rice and corn crops. This study proposes deep learning and transfer learning for diagnosing plant leaf diseases in various types of plants and unstructured imaging environments. To address these challenges, a selection of VGGNet, ResNet50, InceptionV3, and EfficientNetB0 methods was conducted by testing them using laboratory datasets. Based on the testing, the EfficientNetB0 model performed the best. Then, the selected model parameters were tuned, feature extraction and a new dataset was collected in a real-world domain with varying lighting, changing viewpoints and scales, complex backgrounds, similar symptoms between diseases, and occlusion. The results showed that the proposed model performed very well and robustly, with 98% accuracy and a weighted average F1-score of 98% in identifying food crop diseases: blight, rust, blast, blight, tungro, and healthy leaves. This performance indicates that the developed model is highly reliable in classifying leaf diseases in rice and corn. This model is expected to be applied to precision agriculture technology so that farmers can take timely action regarding treatment without further delay.

Author 1: Khoerul Anwar

Author 2: Tubagus Mohammad Akhriza

Author 3: Mahmud Yunus

Keywords: Plant diseases; deep learning; transfer learning; multi-food crops; precision agriculture

Paper 56: InfoCore: AI Driven Named Entity Deduplication and Event Categorization

Abstract: The exponential growth of digital information presents critical challenges for efficient data management, as conventional manual curation methods remain slow, error-prone, and unable to adapt to evolving data streams. This paper presents InfoCore: AI-Driven Entity Deduplication and Event Categorization, an automated framework that leverages artificial intelligence to identify and remove redundant news articles while classifying them by event. Focusing on the political news domain, the system integrates Natural Language Processing, machine learning, and clustering techniques to enhance information retrieval and reduce redundancy. News content is collected via the Newspaper3k library and processed through tokenization, normalization, and entity extraction. Transformer-based models enable named entity recognition, while LLaMA-based large language models, TensorFlow, and PyTorch support text classification and event categorization. Empirical evaluation demonstrates InfoCore’s capacity to detect duplicates and achieve precise event classification with high scalability. The paper contributes a domain-independent architecture for automated data curation and a replicable workflow that improves efficiency and accuracy in large-scale information systems. The results highlight InfoCore’s potential to advance data management practices and inform the design of intelligent, scalable frameworks for handling unstructured digital content.

Author 1: Rohail Qamar

Author 2: Raheela Asif

Author 3: Abdul Karim Kazi

Author 4: Muhammad Ali

Author 5: Muhammad Mustafa

Keywords: Data deduplication; context-aware; event categorization; NLP; large language model

Paper 57: Two-Level Hierarchical Adaptive Dynamic Fusion for CNN–LSTM Integration in Fatigue Level Prediction

Abstract: Driver fatigue is a major contributor to traffic accidents, yet most existing detection systems rely on unimodal inputs or static fusion mechanisms that lack robustness under poor lighting, partially obscured faces, and missing sensor data. This study aims to overcome these limitations by proposing a Hierarchical Adaptive Dynamic Fusion (HADF) model. HADF integrates a two-level adaptive fusion mechanism combining a CNN (ResNet-18) for facial micro-expressions and an LSTM for physiological signals (heart rate, temperature, and accelerometer). The first stage computes adaptive intra-modality weights (α), while the second stage assigns inter-modality weights (γ), enabling context-aware and resilient multimodal integration even under missing-modality conditions. Experiments on a multimodal fatigue dataset show that HADF achieves a validation accuracy of 96.5%, a macro F1-score of 0.96, and ROC-AUC values of 1.00 (Normal), 0.99 (Eye-Closed), and 0.93 (Yawn). Compared with unimodal and static-fusion baselines, HADF improves accuracy by approximately 4.5% and macro F1-score by 6–9%, while maintaining stable performance under incomplete data. These results confirm the novelty of HADF as a two-stage adaptive fusion strategy that enhances accuracy and system robustness, making it suitable for real-time fatigue monitoring in transportation, occupational safety, and healthcare applications.

Author 1: Marlince NK Nababan

Author 2: Poltak Sihombing

Author 3: Erna Nababan

Author 4: T Henny Febriana Harum

Keywords: Multimodal fusion; adaptive dynamic fusion; CNN-LSTM; fatigue level prediction

Paper 58: Optimizer Algorithms Analysis for Intrusion Detection System on Deep Neural Network

Abstract: Intrusion Detection Systems (IDS) play a critical role in identifying potential threats and intrusions in real-time within information technology infrastructures. The development of IDS using Deep Neural Networks (DNN) with the UNSW-NB15 dataset has shown significant potential in improving attack classification accuracy. However, the performance of the DNN-based IDS models is highly dependent on the choice of optimization algorithm. This study compares the performance of several commonly used optimizers in DNN training, including SGD, RMSprop, Adam, Adadelta, Adagrad, Adamax, Adafactor, and Nadam. The quantitative analysis demonstrates that Adam achieves the highest accuracy among all optimizers tested, while Adadelta performs the worst. RMSprop shows instability in both validation accuracy and loss convergence, indicating challenges in adapting the learning rate for consistent learning. The ANOVA analysis yields an F-statistic of 34.687, which is greater than the F-critical value of 2.140 at a significance level of α = 0.05. This result confirms a statistically significant difference in performance among the tested optimization algorithms. These findings provide valuable insights for selecting the most appropriate optimizer to enhance the performance of DNN-based intrusion detection systems. Furthermore, this research contributes to the existing literature by offering a comprehensive comparative evaluation of optimizers, supporting future studies in improving IDS optimization strategies.

Author 1: H. A. Danang Rimbawa

Author 2: Agung Nugroho

Author 3: Muhammad Abditya Arghanie

Keywords: Deep Neural Networks (DNN); Intrusion Detection System (IDS); optimization algorithms; UNSW-NB15 dataset

Paper 59: Shadow IT Transformation in the Post-Pandemic Digital Workplace: A Systematic Literature Review

Abstract: The COVID-19 pandemic has significantly altered organizational work patterns, accelerating digital transformation and the adoption of remote and hybrid work models. These changes have affected the practice of shadow IT, the use of unauthorized IT by employees without formal IT approval. This systematic literature review aims to explore how the pandemic and the shift to remote work have impacted shadow IT adoption, motivations, and management strategies in the context of digital transformation. We followed the PRISMA 2020 guidelines to conduct a search of peer-reviewed articles published between 2018 and 2025 across multiple databases (Scopus, Web of Science, IEEE Xplore, ACM Digital Library, AIS eLibrary). A total of 67 studies were included based on predefined criteria. The review identified key themes related to the evolving nature of shadow IT adoption, its associated risks, and adaptive management practices. Shadow IT adoption increased from 30 to 40% before the pandemic to 41% in 2022, with projections suggesting it could reach 75% by 2027. The findings show a shift in motivation for adopting shadow IT, from convenience-driven use to a necessity for business continuity, and finally, to a strategy for optimizing organizational processes. This review highlights the need for organizations to rethink IT governance in the post-pandemic digital workplace, as shadow IT has moved from an issue to be eliminated to a phenomenon that can be managed and leveraged.

Author 1: Ginanjar Nugraha

Author 2: Munir

Author 3: Puspo Dewi Dirgantari

Keywords: Shadow IT; COVID-19; remote work; digital transformation; systematic literature review; IT governance

Paper 60: HeritageLM: Culturally-Aware Multimodal Language Modeling with Memory-Enhanced Cross-Dialect Adaptation

Abstract: The HeritageLM system solves the acute problem of language loss by proposing a multimodal language model that brings into Generative Memory the cultural context. Manual documentation and linguistic archiving are the traditional ways of preserving dialects, which may not be effective in preserving the phonetic variety and other cultural peculiarities of the face of an endangered dialect. The current NLP models, such as BERT and GPT, are not effective to produce dialectal content because they do not have exposure to under-resourced and historically rich language varieties. These shortcomings are mitigated by training Cultural Contextual Embeddings (CCE), Generative Memory Augmentation (GMA), and Cross-Dialect Contrastive Transfer Learning (CDCP) using reinforcement learning with Cultural Rewards (RLCR). It is a step-by-step process that builds a Multimodal Cultural Knowledge Graph (MCKG), matches dialect embeddings in contrastive learning, and retrieves culturally relevant information in the generation process. The model was trained on the Indian Languages Audio Dataset of Kaggle, which also included phonetic variations of ten languages with preprocessing steps of text-to-speech analysis, phonetic annotation, and semantic tagging. HeritageLM, which was implemented in Python, scored above 98 in its BLEU, ROUGE-L, phonetic accuracy, and cultural embedding, showing that it can effectively generate linguistically accurate, phonetically accurate, and culturally authentic results. These outcomes are a major step towards the resurrection of dying dialects and maintaining their distinct cultural background.

Author 1: Pasupuleti Venkata Ramana

Author 2: K. K. Sunalini

Author 3: A. Swathi

Author 4: R. Lakshmi

Author 5: Raman Kumar

Author 6: Elangovan Muniyandy

Author 7: Khaled Bedair

Keywords: Cultural embeddings; Generative Memory; dialect revitalization; contrastive learning; multimodal NLP

Paper 61: Personalized Grammar Refinement Using Meta-Reinforcement Learning and Transformer-Based Framework

Abstract: Writing competence is an essential academic and professional proficiency, and grammatical precision and reliability a long-term issue, especially among ESL students. Conventional rule-based and statistical grammar correction models have constraints based on context, whereas contemporary Transformer-based sequence-to-sequence models like BERT, T5, and GPT have strong performance but cannot be customized or adapted to specific writer styles. To fill in these gaps, this study introduces Meta-ACGR, a meta-reinforcement learning grammar refinement system that augments Transformer-based seq2seq models with Proximal Policy Optimization (PPO) and Model-Agnostic Meta-Learning (MAML) and curriculum learning. The model promotes individualized grammar correction, which allows quick adjustment to the new learners in ESL by using meta-learning and guided error development. Meta-ACGR is written in Python with the help of PyTorch and trained on big datasets of ESL language, like NUCLE and Lang-8, which can be refined based on context and individual learners. Empirical evidence indicates that Meta-ACGR receives better grammatical accuracy (86.2 vs. 94.0 per cent), decreases inference latency (12 per cent vs. baseline Transformer models), and performs better on personalization (15 per cent vs. baseline Transformer models). Altogether, Meta-ACGR provides a scalable, adaptable, and customized grammar check system with good chances to be implemented in real-life to improve writing in ESL.

Author 1: Bukka Shobharani

Author 2: Melito D. Mayormente

Author 3: Edgardo B. Sario

Author 4: Bernadette R. Gumpal

Author 5: S. Farhad

Author 6: Jasgurpreet Singh Chohan

Author 7: Elangovan Muniyandy

Keywords: Grammar correction; Transformer models; Meta-Reinforcement Learning; curriculum learning; personalization; ESL writing

Paper 62: Graph-Enhanced Transformer Framework for Context-Sensitive English Skill Assessment

Abstract: The integration of Artificial Intelligence (AI) into English Language Teaching (ELT) has enabled personalized and interactive learning, yet most existing systems rely on static, rule-based feedback models, which fail to capture learner history or adapt interventions based on skill interdependencies. These limitations result in generic management, reduced learner engagement, and fragmented skill development. To overcome these challenges, this study proposes a hybrid DeBERTa–GAT–PPO framework that combines transformer-based contextual embeddings, graph attention-based inter-skill modeling, and reinforcement learning for adaptive, history-aware feedback. The model is implemented in Python 3.10 using PyTorch 2.0 and processes the Kaggle Feedback Prize – English Language Learning dataset, containing over 6,600 annotated essays across cohesion, syntax, vocabulary, phraseology, grammar, and conventions. Learner essays are preprocessed, embedded via DeBERTa, and represented as a knowledge graph to capture skill interdependencies through GAT. The PPO agent then generates context-sensitive feedback optimized via policy gradients. Experimental results demonstrate that the proposed framework achieves an accuracy of 89.8% and an AUC of 0.96, representing an approximate 6 to 8% improvement over baseline models such as BERT and RoBERTa. Visualizations and ablation studies confirm effective learning of inter-skill dependencies and reinforcement-based feedback adaptation. Overall, the proposed model provides scalable, interpretable, and pedagogically effective feedback, bridging the gap between conventional AI tutors and fully adaptive, learner-centered systems, thus advancing the state-of-the-art in intelligent English language tutoring.

Author 1: Anna Shalini

Author 2: Myagmarsuren Orosoo

Author 3: W. Grace Shanthi

Author 4: Prema S

Author 5: S. Farhad

Author 6: Elangovan Muniyandy

Author 7: A. Chrispin Antonieta Dhivya

Keywords: Memory-augmented networks; conversational AI; English Language Teaching (ELT); adaptive feedback; personalized language learning

Paper 63: Attention-Enhanced Multi-View Graph Convolutional Network for Early Prediction of Chronic Kidney Disease

Abstract: The prediction of chronic kidney disease (CKD) must have models capable of processing heterogeneous clinical data and being transparent to assist clinical decision making. Current CKD research usually uses single-view data, integrated graph representations, or bivalent deep learning systems that do not reflect view-specific clinical connections or cannot be interpreted effectively. The first study that uses a combination of the individual multi-view similarity graphs and an attention-based fusion approach to predict the risk of CKD, and the study overcomes the shortcomings of the earlier machine learning, deep learning, and graph-based models. The suggested Attentive Multi-View Graph Convolutional Network (MV-GCN-Attn) uses Graph Convolutional Networks to learn view-specific embeddings and applies them in an adaptive way with the help of attention mechanisms and highlighted clinically influential features. The model has an accuracy of 91.0% along with a precision of 89.0% and a recall of 92.0% and F1-score of 90.0 in the experiment of 400 patient records and 24 attributes in a publicly available dataset of UCI CKD, which is higher than the conventional baselines. The framework also offers feature- and view-level interpretability and the key indicators are determined: serum creatinine and haemoglobin. These results indicate that the use of multi-view graph learning with attention-based interpretability has the potential to provide effective, clinically significant predictions, which can be used with a high degree of confidence in the practical implementation of CKD screening and decision-support in the work of various healthcare facilities and as a valuable aid in the early clinical intervention process.

Author 1: Roshan D Suvaris

Author 2: K Nagaiah

Author 3: P. Satish

Author 4: Hussana Johar R B

Author 5: Elangovan Muniyandy

Author 6: Manasa Adusumilli

Author 7: Khaled Bedair

Keywords: Chronic kidney disease progression; multi-view graph convolutional network; temporal fusion transformer; uncertainty-aware AI models; personalized medicine in healthcare

Paper 64: Multimodal Cognitive Mapping Framework for Context-Aware Figurative Language Understanding

Abstract: Learning figurative language, including idioms, metaphors, and similes, remains challenging due to subtle cultural, contextual, and multimodal cues that cannot be inferred from literal meanings alone. Traditional unimodal and text-only approaches, such as CLS-BERT, LaBSE, and mUSE, often fail to capture these deeper semantic patterns, resulting in reduced accuracy and limited cultural generalization. This study introduces a context-aware multimodal learning framework that integrates textual embeddings from a Graph-Enhanced Transformer (HCGT) with visual embeddings from CLIP, fused through a graph-based cross-modal attention mechanism, and refined using a cognitive mapping layer. This architecture models human-like semantic reasoning by aligning literal and figurative senses across modalities while maintaining conceptual structure through graph-driven representation learning. Experiments conducted on idiom, metaphor, simile, and multimodal meme datasets include preprocessing steps such as text cleaning, tokenization, image normalization, and label standardization. The framework achieves an accuracy of 90%, surpassing state-of-the-art text-only transformer baselines by 3–4%. Explainable AI tools, including attention heatmaps and SHAP values, validate the interpretability of the model by highlighting influential textual tokens and visual regions. The results confirm that integrating multimodal embeddings with cognitive mapping substantially enhances performance, interpretability, and cultural sensitivity in figurative language understanding.

Author 1: R. Swathi Gudipati

Author 2: Neena PC

Author 3: K. Ezhilmathi

Author 4: M. Durairaj

Author 5: S. Farhad

Author 6: Elangovan Muniyandy

Author 7: Padmashree V

Keywords: Bi-LSTM; cognitive mapping; cross-lingual understanding; idiom acquisition; multimodal learning

Paper 65: Unveiling Gender in Malay-English Short Text: A Comparative Study of ML, DL and Sequential Models with XAI Misclassification Analysis

Abstract: Gender identification through written text analysis leverages writer-specific characteristics including linguistic patterns and stylistic behaviors, yet research on gender identification in Malay-English (Manglish) using Traditional Machine Learning (ML), Shallow Deep Learning (DL), and Deep Sequential techniques remains limited compared to English-focused studies. This study addresses this gap by investigating gender identification in Manglish across traditional ML, Shallow DL, and Sequential Deep Model approaches using a self-collected dataset of Manglish tweets from 50 anonymized Malaysian public figures. Following preprocessing, feature extraction employed Word2Vec embeddings and TF-IDF methods, revealing that Word2Vec embeddings delivered superior performance across Shallow DL and Deep Sequential models, with Bi-CNN achieving optimal results of accuracy (0.722), precision (0.727), recall (0.722), and F1-score (0.720), while TF-IDF vectorization yielded substandard performance except for Logistic Regression, which achieved consistent metrics of 0.728 across all evaluation criteria. To enhance model interpretability, eXplainable Artificial Intelligence (XAI) tools including SHAP and LIME were applied to analyze misclassifications, identifying key issues such as frequent shortform usage and word misassignment affecting prediction accuracy, and incorporating these XAI insights through iterative refinements yielded modest improvements from 72.4% to 72.8%, demonstrating XAI's value in model optimization despite limitations in capturing dataset biases and complex linguistic patterns. This study contributes the first gender classification dataset for Malay short text and demonstrates that Shallow DL and Deep Sequential models, enhanced by XAI-driven analysis, show significant promise for mixed-language contexts, highlighting the unique challenges of code-switched languages in NLP tasks while suggesting future research should explore large language models to advance classification performance in multilingual social media environments.

Author 1: Norazlina Khamis

Author 2: Nur Shaheera Shastera Nulizairos

Author 3: Haslizatul Mohamed Hanum

Author 4: Amirah Ahmad

Author 5: Nor Hapiza Mohd Ariffin

Author 6: Ruhaila Maskat

Keywords: Gender identification; Manglish; machine learning; shallow deep learning; deep sequential model

Paper 66: Causality Aware Multimodal Reasoning Network in Human Emotion Identification and Sentiment Understanding

Abstract: Sentiment and emotion recognition in dynamic English communication require intelligent systems capable of reasoning beyond surface correlations among linguistic, acoustic, and visual cues. Traditional multimodal approaches exhibit limited interpretability, weak contextual adaptability, and lack causal understanding of emotional expressions, resulting in inconsistent predictions under ambiguous conditions. To address these challenges, a Context-Adaptive Knowledge-Guided Causal Reasoning Network (CKCR-Net) is introduced, integrating external semantic and affective knowledge with multimodal fusion to ensure transparency and contextual sensitivity. The proposed framework employs a Dynamic Multimodal Knowledge Graph (DMKG), hierarchical cross-modal attention, and a dual-stage causal reasoning module to infer cause–effect dependencies among modalities. The model was implemented in Python (PyTorch) using the CMU-MOSEI benchmark dataset and optimized through Adam optimizer and consistency-based loss regularization. CKCR-Net achieved an accuracy of 97.5%, precision of 96.4%, recall of 97.2%, and F1-score of 97.3%, significantly outperforming models such as CM-BERT (89.4%), RoBERTa (71%), and TFIDF-based fusion (96.9%). The causal reasoning mechanism improved recognition of subtle emotions like sarcasm and empathy, enhancing interpretability through attention heatmaps and counterfactual analysis. Overall, CKCR-Net provides an explainable, context-sensitive, and high-performing framework for multimodal sentiment analysis, offering a reliable pathway toward transparent affective computing and human–machine communication.

Author 1: N. K. Thakre

Author 2: Yazan Shaker Almahammed

Author 3: G. Indra Navaroj

Author 4: Mohammed Fahad Almohazie

Author 5: Abdullah Albalawi

Author 6: Marran Al Qwaid

Author 7: G. Sanjiv Rao

Keywords: Multimodal sentiment analysis; knowledge-driven transformer; explainable AI; dynamic multimodal fusion; CMU-MOSEI dataset

Paper 67: Beyond Ensembles: Architecture-Level Fusion for Enhanced Monument Heritage Recognition

Abstract: Heritage is seen as a key part of nations, including a broad variety of traditions, cultures, monuments, plants and animals, foods, music, and further. Regarding countries, their own heritages are defined by preservation, excavation, and restoration of historical assets that are important and show the nation's history. It comprises a wide range of physical objects and materials found in cultural institutions which are moveable heritage, as well as the heritage found in built environments which are immovable and natural landscapes. Previous studies on monument classification frequently used single small datasets, limiting accuracy and generalizability. This work introduces a proposed model and a thorough experimental comparison of widely used deep learning architectures, specifically Convolutional Neural Networks and Transformers beside our proposed model, for monument recognition in the cultural monument domain. It seeks to conduct a comparative experiment for selecting representatives from these two methodologies regarding their capacity for transferring information from a general dataset, like ImageNet, to heritage landmarks datasets of varying sizes. When we tested samples of the topologies ResNet, DenseNet, and Swin Transformer (Swin-T), we find that the proposed model had the best results, however ResNet-50 achieved comparable accuracy to Swin-T.

Author 1: Mennat Allah Hassan

Author 2: Mona M. Nasr

Author 3: Alaa Mahmoud Hamdy

Keywords: Cultural monument; heritage landmarks; monument classification; monument recognition; transformers

Paper 68: Semi-Supervised Learning vs. Few-Shot Learning: Which is Better for Sentiment Analysis on Hotel Reviews Towards a Small Labeled Training Data?

Abstract: The massive growth in user reviews on the online travel agent (OTA) website can be automatically processed using sentiment analysis to understand consumer satisfaction or feedback. Sentiment analysis is commonly implemented as a sentiment classification task by applying classical machine learning and deep learning algorithms. However, implementing both strategies has a significant challenge in providing a reliable labeled dataset since labeling is time-consuming and highly resource-intensive. Therefore, this study aims to compare the performance of two learning methods: semi-supervised learning (SSL) and few-shot learning (FSL), since there is still no direct, controlled comparison between both methods. SSL is a learning method that builds a generalized classification model as a refined model based on automatically generated additional labeled data. In contrast, FSL is a learning method that enables a generalized pre-trained model to predict unlabeled data using only a few labeled samples per class. This study evaluates the self-training method on SSL, and the implemented FSL algorithm is Sentence Transformer Fine-Tuning (SETFIT). The results show that implementing FSL (employing only 16 labeled training samples) outperforms SSL with an accuracy improvement of 9.5%. The implementation of SETFIT is very promising as a solution to overcome the limited amount of labeled data in the classification task. Moreover, SETFIT is more adaptable to various low-resource language domains than other, more data-intensive learning approaches.

Author 1: Retno Kusumaningrum

Author 2: Ahmad Ainun Herlambang

Author 3: Wafiq Afifah

Author 4: Adi Wibowo

Author 5: Sutikno

Author 6: Priyo Sidik Sasongko

Keywords: Sentiment analysis; hotel reviews; semi-supervised learning; few-shot learning; low-resource language

Paper 69: Linguistically Informed Essay Assessment Framework to Analyze Writing Style Vocabulary Usage and Coherence

Abstract: Automated essay scoring (AES) has become an essential tool in educational technology, yet many existing approaches rely on black-box models that lack interpretability and adaptability across diverse prompts and writing styles. Conventional transformer-based AES systems demonstrate strong accuracy, but often fail to provide pedagogically meaningful feedback or generalize effectively in low-resource settings, limiting their practical applicability. The proposed COSMET-Net (Contrastive and Explainable Semantic Meta-Evaluation Network) addresses these limitations by integrating contrastive learning, meta-learning, and explainable AI to produce an adaptive and interpretable evaluation of academic essays. Essays are processed through text cleaning, tokenization, and lemmatization, and embeddings are generated using pretrained transformers such as BERT and RoBERTa. Contrastive learning distinguishes high- and low-quality essays, while a Contrastive Linguistic Regularization (CLR) layer aligns embeddings with linguistic properties, enhancing interpretability. Meta-learning enables rapid adaptation to novel prompts with minimal additional data. The explainable output module, employing attention visualization and SHAP values, provides detailed feedback on grammar, coherence, vocabulary richness, and readability. The framework was implemented in Python with PyTorch and Hugging Face Transformers and evaluated on the IELTS Writing Scored Essays Dataset. COSMET-Net achieved an accuracy of 92%, a recall of 93%, and an F1-score of 92%, surpassing existing models such as hybrid RoBERTa + linguistic features (F1-score 84%) and discourse + lexical regression (F1-score 88%). These results demonstrate that COSMET-Net delivers highly accurate, flexible, and linguistically interpretable assessments, providing a scalable solution for automated and pedagogically meaningful essay evaluation.

Author 1: Sreela B

Author 2: B. Neelambaram

Author 3: Manasa Adusumilli

Author 4: Revati Ramrao Rautrao

Author 5: Aseel Smerat

Author 6: Myagmarsuren Orosoo

Author 7: A. Swathi

Keywords: COSMET-Net; contrastive learning; explainable AI; meta-learning; essay scoring

Paper 70: An Interpretable Dual-Level Feedback Approach for Improving Graded Language Simplification and Readability

Abstract: Text simplification plays a vital role in adaptive language learning, especially when aligned with the Common European Framework of Reference (CEFR) proficiency levels. The purpose of this study is to develop an interpretable and CEFR-aligned text simplification framework that produces pedagogically appropriate simplified texts for learners at different proficiency levels. Existing neural simplification approaches, such as ACCESS, MUSS, and EditNTS, primarily rely on single-level feedback or surface-level readability measures, limiting their ability to ensure both sentence-level linguistic simplicity and document-level coherence. To address these gaps, this study proposes CEFR-RefineNet, a hybrid framework integrating T5 for generative simplification and BERT for contextual CEFR-level classification, enhanced through a novel Dual-Level Explainable Feedback Loop (DL-EFL). The DL-EFL simultaneously evaluates sentence-level linguistic difficulty and document-level readability while providing token-level error attribution for interpretability. The model was implemented using Python and the Hugging Face Transformers library, trained and tested on the CEFR Levelled English Texts corpus comprising 1,500 texts spanning levels A1 to C2. Experimental results show that CEFR-RefineNet achieved a SARI score of 0.78, accuracy of 91%, and F1-score of 0.85, outperforming the strongest baseline (MUSS, 81% accuracy) by approximately 12%. The adaptive feedback mechanism accelerated reward convergence and improved CEFR compliance, ensuring more pedagogically suitable simplifications. In summary, the proposed CEFR-RefineNet establishes a transparent, interpretable, and performance-driven text simplification model capable of generating fluent, meaning-preserving, and CEFR-aligned texts, paving the way for intelligent and adaptive language-learning systems.

Author 1: Pavani G

Author 2: Myagmarsuren Orosoo

Author 3: W. Grace Shanthi

Author 4: Vinod Waiker

Author 5: Aseel Smerat

Author 6: Bhuvaneswari Pagidipati

Author 7: Bansode G. S

Author 8: Osama R. Shahin

Keywords: Text simplification; CEFR-level alignment; reinforcement learning; adaptive feedback loop; T5 transformer

Paper 71: Meta Learning Enhanced Graph Transformer for Robust Smart Grid Anomaly Detection

Abstract: The increasing complexity of modern smart grids and the heterogeneity of multi-sensor data make anomaly detection extremely challenging, as existing techniques struggle to capture long-range spatial dependencies, cross-sensor interactions, and unseen anomaly patterns. Conventional models such as Isolation Forest, Random Forest, GCAD, AT-GTL, CVTGAD, and hybrid CNN-Transformer approaches often suffer from limited generalization, weak multimodal fusion, and strong dependence on labeled anomalies. To address these limitations, this study introduces a novel Multimodal Graph Transformer with Contrastive Self-Supervised Learning and Model-Agnostic Meta-Learning (MGT-CGSSML), a uniquely integrated framework designed to learn structural, attribute, and cross-modal relationships simultaneously. The proposed method stands out by combining multimodal graph encoding, dual-view contrastive learning, and fast meta-adaptation, enabling the model to rapidly identify new anomaly types with minimal labeled data. Implemented in Python using PyTorch, the model is evaluated on a multimodal smart grid dataset containing time-stamped voltage, current, power factor, frequency, temperature, and humidity measurements recorded at 15-minute intervals. Experimental results demonstrate 96.5% accuracy, 95% precision, 95.5% recall, and 95.2% F1-score, reflecting a 3–5% performance improvement over advanced baseline models due to enhanced multimodal fusion and meta-learning optimization. The study concludes that MGT-CGSSML delivers a scalable, interpretable, and real-time anomaly detection solution capable of supporting resilient and adaptive smart-grid operations, offering substantial advancements over existing methods.

Author 1: Layth Almahadeen

Author 2: Aseel Smerat

Author 3: Sandeep Kumar Mathariya

Author 4: G. Indra Navaroj

Author 5: Vuda Sreenivasa Rao

Author 6: Kamila Ibragimova

Author 7: Osama R.Shahin

Keywords: Adaptive detection; anomaly detection; contrastive learning; graph transformer networks; smart grid

Paper 72: An Interpretable Deep Learning Framework for Measuring Organizational Digital Transformation Readiness

Abstract: The accelerating pace of digital transformation (DT) across industries demands accurate, transparent, and adaptable maturity evaluation frameworks capable of capturing complex organizational behaviors. Conventional fuzzy logic and decision tree-based maturity models cannot effectively represent the nonlinear dependencies among DT indicators and often produce inconsistent, opaque assessments. To overcome these limitations, this study proposes the TUMI (Transformer TabNet Unified Maturity Intelligence) framework, a novel hybrid deep learning architecture specifically designed for DT maturity assessment. The framework uniquely integrates FT-Transformer and TabNet, enabling simultaneous modeling of global feature dependencies through attention mechanisms and localized sparse feature selection aligned with DT maturity metrics. This domain-tailored hybridization goes beyond existing hybrid or ensemble approaches by supporting real-time readiness estimation, accommodating heterogeneous organizational indicators, and offering structured interpretability based on complementary attention weights and feature selection masks. The proposed model was trained using a multi-dimensional DT maturity dataset implemented in Python (PyTorch). Experimental results demonstrate strong predictive performance, with 97.0% accuracy, 96.0% precision, 95.0% recall, and an AUC of 98.2%, representing an 8.5% improvement over traditional fuzzy and decision tree models. The interpretability provided by the combined mechanisms offers clearer insight into the organizational determinants influencing maturity progression. Overall, TUMI enhances transparency, diagnostic capability, and scalability, providing an evidence-based, explainable, and cross-industry applicable solution for supporting organizations in evaluating and improving their digital transformation maturity.

Author 1: Pravin D. Sawant

Author 2: Veera Ankalu Vuyyuru

Author 3: B. Arunsundar

Author 4: A. Vini Infanta

Author 5: Dekhkonov Burkhon

Author 6: N. Roopalatha

Keywords: Digital transformation; FT-Transformer; TabNet; maturity intelligence; deep learning

Paper 73: Clinically Informed Adaptive Multimodal Graph Learning Paradigm for Transparent Temporal and Generalizable Alzheimer’s Disease Diagnosis

Abstract: This is a clinically reliable and explainable diagnostic framework for the early detection of Alzheimer's disease with multimodal data. Current computational methods face challenges in dealing with fragmented clinical information, poor cross-modal integration, limited temporal modelling, and low interpretability, rendering them unsuitable for real-world medical deployment. To overcome these limitations, we propose the Clinically Guided Adaptive Multimodal Graph Transformer (CAM-GT), a novel architecture that fuses clinical priors with graph-based learning and transformer-driven temporal reasoning within a unified model. The proposed framework uniquely integrates clinically guided graph attention, cross-modal fusion, and contrastive alignment, where the system can capture hidden relationships among imaging, cognitive scores, and clinical biomarkers with high robustness against missing or imbalanced modalities. Implemented on the Python platform with advanced deep-learning libraries, CAM-GT carries out multimodal encoding, temporal progression modeling, and explainability mapping in order to identify the most significant biomarkers that influence the status of a disease. Experimental evaluation demonstrates that the model performs well by achieving an accuracy of 97%, a 97.2% AUC, and outperforming existing models while maintaining strong generalization in heterogeneous clinical environments. Further, high interpretability ensures that clinically, it will be able to trace how predictions are made to instill greater trust and ethical reliability and increase the adoption potential in hospitals and research centers. Finally, CAM-GT benefits neurologists, radiologists, healthcare institutions, and researchers by providing a stable, transparent, high-performing AI system that has the capability to support early diagnosis and guide real-world clinical decision-making in neurodegenerative disease care.

Author 1: Padmavati Shrivastava

Author 2: V S Krushnasamy

Author 3: Guru Basava Aradhya S

Author 4: Vinod Waiker

Author 5: Peddireddy Veera Venkateswara Rao

Author 6: Elangovan Muniyandy

Author 7: Khaled Bedair

Keywords: Alzheimer’s detection; graph neural network; multimodal fusion; explainable AI; temporal transformer

Paper 74: Transformer Driven Multi-Agent Reinforcement Learning Framework for Integrated Waste Classification Forecasting and Adaptive Routing

Abstract: The rapid expansion of urban populations has intensified the challenges associated with municipal solid waste management, particularly where conventional static or ad-hoc routing strategies create operational inefficiencies, excessive fuel usage, and repeated bin overflow. Many existing systems still treat waste classification, fill-level forecasting, and routing as separate processes, which restricts coordinated optimization and limits broader sustainability outcomes. To address these shortcomings, TMORL is introduced as a Transformer-enhanced Multi-Agent Reinforcement Learning framework that unifies perception, prediction, and decision-making for intelligent waste management. The framework integrates IoT-enabled sensor measurements with deep learning and MARL-driven optimization to manage waste collection adaptively under real-time uncertainty. A Vision Transformer supports precise waste image classification through global spatial feature extraction, while a Temporal Fusion Transformer generates accurate, uncertainty-aware multi-horizon fill-level forecasts. These model outputs collectively shape the state representation for a multi-objective MARL module that optimizes fuel consumption, travel duration, emission reduction, and overflow mitigation, enabling simultaneous operational and sustainability improvements. TMORL is implemented in PyTorch and evaluated using the Smart Waste Management Dataset containing heterogeneous IoT bin measurements and annotated waste images. The model achieves strong perception accuracy, reporting 97.3% precision, 96.6% recall, and 98.4% mAP@0.5, while the TFT forecasts align closely with real bin-fill patterns to support proactive routing adjustments. When compared with static scheduling and Ant Colony Optimization routing, TMORL reduces fuel usage by 22%, collection time by 25%, and overflow incidents by 95%. Overall, the findings confirm that a transformer-driven, IoT-integrated MARL framework significantly strengthens efficiency, decision responsiveness, and environmental sustainability in next-generation smart waste management systems.

Author 1: Ritesh Patel

Author 2: Igamberdiyev Asqar Kimsanovich

Author 3: Vinod Waiker

Author 4: Elangovan Muniyandy

Author 5: Swarna Mahesh Naidu

Author 6: Nurilla Mahamatov

Author 7: Osama R.Shahin

Keywords: Smart waste management; temporal fusion transformer; vision transformer; predictive analytics; route optimization; deep learning

Paper 75: Leveraging Intelligent Speech Training to Elevate Phonetic Accuracy and Prosodic Fluency in English Learners

Abstract: The successful teaching of pronunciation, as well as prosody, is the significant challenge that still remains to the English as Foreign Learning (EFL) students. Traditional pedagogical theories tend to focus on segmental phoneme accuracy but ignore suprasegmental components (stress or rhythm and intonation) which are natural and intelligible speech components. The currently available systems of computer-assisted pronunciation training (CAPT) are useful, but limited by the fact that they are based on limited acoustic models and incomplete coverage of prosodic characteristics, leading to less than optimal accuracy and limited pedagogical suitability. To overcome these shortcomings, the current paper proposes Attention-Guided Cross-Lingual Self-Supervised Learning (AG-CLSSL), a new model that is both able to combine phoneme-level representations of XLS-R (wav2vec2-large-xlsr-53) and prosodic representations of the pitch, energy, and duration through a Phoneme-Prosody Cross-Attention Fusion (PP-CAF) process. This conglomeration allows the joint and context specific representation of the speech that is further refined by the multi-task Transformer-based scoring model to jointly assess the accuracy of pronunciation, the consistency of the prosody and the general intelligibility. The framework is implemented in Python, with support of PyTorch and Hugging Face Transformers and is trained on an evaluated corpus of EFL learner speech (n=100) with a variety of L1 backgrounds, including Mandarin, Hindi, and Spanish. Experimental assessments indicate significant performance improvement with 55.4% decrease in Phoneme Error rate, 52.0 percent decrease in Word Error rate, 43.3 percent increase in Stress Placement Accuracy and 34.9 percent increase in Pitch Alignment Score. The total acoustic similarity to native speech went up by 36.1, which demonstrates the ability of AG-CLSSL to progress articulatory accuracy as well as the naturalness of prosody and provide interpretable and attention-directed information on scalable AI-based pronunciation and prosody training.

Author 1: Amit Khapekar

Author 2: Nidhi Mishra

Author 3: Vijaya Lakshmi Mandava

Author 4: T K Rama Krishna Rao

Author 5: Bhuvaneswari Pagidipati

Author 6: Prasad Devarasetty

Author 7: Elangovan Muniyandy

Keywords: Automatic speech recognition; pronunciation and prosody; transformer-based phoneme identification; prosody assessment; adaptive learning algorithm

Paper 76: An Enhanced Deep Learning Framework for Diabetic Retinopathy Classification Using Multiple Convolutional Neural Network Architectures

Abstract: Diabetic retinopathy (DR) is a leading cause of blindness, requiring early and accurate diagnosis. Although deep learning, particularly Convolutional Neural Networks (CNNs), has shown promising results in automating DR classification, selecting the optimal architecture and extracting effective features for specific clinical datasets remains a challenge. This study aims to conduct a comprehensive performance evaluation of six CNN architectures—DenseNet121, MobileNet, NasNet_Mobile, ResNet50, VGG16, and VGG19—for DR classification on a dataset from the Community Eye Hospital of South Sumatra Province. The main novelty of our approach lies in a specific preprocessing workflow that integrates grayscale conversion and Canny edge detection to enhance the visibility of critical retinal features, such as blood vessels and lesions, before classification. Using a dataset of 3000 fundus images across five classes (No_DR, Mild, Moderate, Severe, and Proliferative DR), the model was trained with data augmentation and the Adam optimizer. Experimental results indicate that the VGG16 architecture achieves a peak accuracy of 73%, outperforming baseline implementations from previous studies. This study highlights the potential of combining classical CNN models with tailored preprocessing for improved DR detection, thus providing a benchmark for model selection on similar clinical datasets. These findings highlight the robustness and stability of VGG16, demonstrating its suitability as an early DR screening tool.

Author 1: Zaid Romegar Mair

Author 2: Agus Harjoko

Author 3: Rendra Gustriansyah

Author 4: Septa Cahyani

Author 5: Rudi Heriansyah

Author 6: Indah Permatasari

Author 7: Muhammad Haviz Irfani

Keywords: Diabetic retinopathy; diabetic retinopathy classification; deep learning; Convolutional Neural Network (CNN); VGG16

Paper 77: Integrating Causality with Spatio-Temporal Attention for Accurate Airline Delay Prediction

Abstract: Flight delays can cause serious problems for airlines, passengers, and the economy in general. Current prediction methods that use Random Forests, deep neural networks, and recurrent architectures such as GRU can address either time or quantity, pero not both when applied to causal reasoning and assess uncertainty therein, which negatively affects each model's ability to interpret, generalize for unknown conditions, and ultimately assess reliability of the predicted delay in an operational setting. Causal-Aware Spatio-Temporal Attention Network (CASTAN) is designed as a combined approach to address these challenges of spatio-temporal and causal modeling all in one. Analysts use GraphSAGE-based spatial encoding to encode and capture inter-airport dependencies, with a self-attention temporal encoder to learn long-range sequential patterns of historical delays in addition to traffic and weather factors. A cross-attention fusion mechanism accounts for the dynamic and spatio-temporal contributions to delay. A final causal counterfactual module adds interpretable independence results—helping analysts to assess the contributing factors to delay. Finally, the incorporation of dropout is done in a Bayesian approach to assess uncertainty for each prediction made and generate uncertainty-aware predictions so analysts may assess reliability through levels of confidence or any other metric decided. Results from evaluation of a large-scale U.S. flight dataset compared to traditional baselines demonstrate the predictive power of the model, achieving 96.4% accuracy, RMSE of 4.2, and MAE of 2.9. The CASTAN process has positioned its place as an interpretable, reliable, and operationally informative modeling approach to proactive management of airline delay.

Author 1: Akash Daulatrao Gedam

Author 2: Pavaimalar S

Author 3: Mercy Toni

Author 4: Y. Rajesh Babu

Author 5: P. Satish

Author 6: Bobonazarov Abdurasul

Author 7: Elangovan Muniyandy

Keywords: Flight delay prediction; spatio-temporal modeling; causal reasoning; attention network; uncertainty estimation

Paper 78: AI-Enabled Assessed Learning from an Interdisciplinary Educational Perspective: An Empirical Case Study Using ChatGPT LLM

Abstract: The recent developments in artificial intelligence (AI) have introduced transformative tools into the educational landscape, with early signs of its impact on advancing student learning methods and possible outcomes. This paper explores the possible integration of large language models (LLMs) in higher education academic courses, using the popular ChatGPT LLM as a case study in interdisciplinary academic courses. In particular, this paper investigates the potential to advance learning and creativity in interdisciplinary courses and their assessment processes. First, we provide a qualitative review of the state-of-the-art in LLMs and chatbot uses in higher education, highlighting some notable gaps in the literature that motivated this work. Next, the paper explores the application of ChatGPT in interdisciplinary courses, examining how its capabilities can be leveraged to support multi-disciplinary learning and collaboration through several assessment types in interdisciplinary courses. The analysis is used to investigate opportunities for fostering deeper integration of knowledge across fields, with LLMs serving as a versatile tool for enhancing student engagement and understanding. Moreover, the paper investigates the role of LLMs in advancing engagement in creative disciplines, exploring how they can be adapted to stimulate creativity and innovation in educational settings. Unique challenges and opportunities presented by incorporating AI tools into creative processes are considered, and their impact on both students and educators is assessed. An empirical design used in this study included testing ChatGPT with 50 test questions across multiple interdisciplinary courses, with 5 evaluators’ perspectives provided using the Delphi scoring method. Key results obtained showed an overall accuracy rate of 93% and 100% for open-ended and multiple-choice questions, respectively. Such promising results had demonstrated ChatGPT’s potential as a constructive tool for fostering interdisciplinary learning by bridging knowledge gaps and promoting integration of ideas across different courses/domains. Finally, this paper aims to contribute to ongoing discussions about the responsible and effective use of AI in academic environments, while highlighting its potential to reform learning and teaching across disciplines.

Author 1: Ali Hassan

Author 2: Omar Tayan

Author 3: Abdurazzag Almiladi

Author 4: Adnan Ahmed Abi Sen

Keywords: Educational assessment; interdisciplinary learning; ChatGPT; Large Language Models (LLMs); Artificial Intelligence (AI)

Paper 79: A Privacy Protection Method for IoT Data Based on Edge Computing and Federated Learning Algorithm

Abstract: This paper proposes a privacy protection method for IoT data integrating edge computing and federated learning. To address challenges including edge node heterogeneity, central server bottlenecks in traditional federated learning, and high overhead of homomorphic encryption, we design a hierarchical architecture comprising requesters, participants, edge nodes, a sensing platform, and a key generation center. Participants train models locally using SGD, encrypt parameters with an optimized verifiable dual-key ElGamal homomorphic encryption scheme, and transmit them to edge nodes. Edge nodes employ the MPSDGS algorithm for participant similarity discovery and dropout supplementation, and the MP-Update method for dynamic weighted averaging to ensure continuity and accuracy. Edge-side ciphertext aggregation reduces data volume to the platform. The sensing platform performs global secure aggregation in ciphertext. Experiments demonstrate that the method maintains data privacy above 0.8, with training and aggregation delays within acceptable ranges for typical IoT scales, balancing privacy and efficiency.

Author 1: Ying Wu

Keywords: Edge computing; federated learning; Internet of Things; privacy protection; homomorphic encryption; dropout supplementation

Paper 80: TOPSIS-YOLO Decision Fusion with Mel-Spectrogram Analysis for Engine Fault Detection

Abstract: Industrial machinery fault detection systems require both high diagnostic accuracy and computational efficiency for real-time deployment. This study presents a novel hybrid approach that integrates the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) with You Only Look Once (YOLO) deep learning for efficient audio-based fault detection in industrial machinery. The proposed methodology employs a two-tiered decision fusion strategy: TOPSIS serves as a rapid mathematical pre-filter analyzing seven acoustic features (RMS, ZCR, Spectral Centroid, Spectral Bandwidth, Peak Frequencies, Kurtosis, and Skewness) extracted from preprocessed 1-2 second audio segments, while YOLO performs detailed spectrogram-based visual analysis on flagged segments. The TOPSIS algorithm normalizes feature vectors, calculates closeness coefficients to ideal and negative-ideal solutions, and classifies segments using a threshold of τ = 0.65. Segments identified as normal terminate processing immediately, while potentially abnormal segments proceed to spectrogram generation and YOLO-based detection. Experimental results on 150 industrial audio segments demonstrate that the hybrid system achieves 93.8% detection accuracy while reducing computational overhead by 85.3% compared to full-dataset YOLO analysis. The TOPSIS pre-filter successfully identifies 128 normal segments (85.3%) with a mean closeness coefficient Ci = 0.847 ± 0.025, while 22 abnormal segments (14.7%) with Ci = 0.084 ± 0.033 are forwarded to YOLO for confirmation. The decision fusion logic enables YOLO to override false positives and flag low-confidence cases for expert review, combining the speed of mathematical analysis with the robustness of deep learning. This approach reduces processing time by approximately 6.8×, decreases GPU utilization by 85%, and minimizes storage requirements for spectrogram images, making it suitable for real-time industrial monitoring systems where computational resources are constrained.

Author 1: Rommel F. Canencia

Author 2: Ken D. Gorro

Author 3: Deolinda E. Caparroso

Author 4: Marlito V. Patunob

Author 5: Jonathan C. Maglasang

Author 6: Joecyn N. Archival

Keywords: Industrial fault detection; audio signal processing; spectrogram analysis; deep learning; multi-criteria decision making (MCDM); TOPSIS; YOLO

Paper 81: Kalman-Enhanced Deep Reinforcement Learning for Noise-Resilient Algorithmic Trading in Volatile Gold Markets

Abstract: Precious metals market, such as gold (XAU/USD), exhibit high volatility and significant microstructure noise in their financial time series, which degrade the reliability of algorithmic trading models. While deep reinforcement learning (DRL) has shown strong results in equities and cryptocurrencies, its application to precious metals remains limited by unstable signals and rapid market fluctuations. This study proposes a Kalman-enhanced DRL framework that integrates classical noise filtering with modern neural architectures to improve signal quality and trading performance in highly volatile environments. The methodology applies Kalman filtering to recursively denoise OHLCV price data, which then serves as an input alongside 22 technical indicators to train three state-of-the-art DRL agents: Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Recurrent PPO (RPPO). Eight years of hourly XAU/USD data (January 2017 to January 2025, N = 47,304) were used for training and evaluation. Models were evaluated on cumulative return, CAGR, Sharpe ratio, maximum drawdown, and volatility. Results demonstrate substantial gains from noise attenuation: PPO with Kalman filtering achieved 80.21% cumulative return (27.1% CAGR, Sharpe 12.10, drawdown -0.48%) compared with raw PPO’s 8.70% (3.46% CAGR, Sharpe 0.45, drawdown - 12.52%). DQN and RPPO achieved comparable improvements, with 244 to 822% return increases, 88 to 96% drawdown reduction, and up to 29× Sharpe ratio enhancement. Statistical significance was confirmed (p < 0.001 for PPO/RPPO; p < 0.05 for DQN). These findings highlight Kalman-enhanced reinforcement learning as a scalable and robust framework for institutional algorithmic trading, bridging signal processing and artificial intelligence for next-generation adaptive trading systems.

Author 1: Amine Kili

Author 2: Brahim Raouyane

Author 3: Mohamed Rachdi

Author 4: Mostafa Bellafkih

Keywords: Deep reinforcement learning; Kalman filtering; algorithmic trading; gold markets; XAU/USD; microstructure noise; signal processing; financial time series; noise reduction; institutional trading

Paper 82: Iterative Partition Optimization: A Novel Approach for Feature Selection in NIR Spectroscopy

Abstract: Machine learning for near-infrared (NIR) spectroscopy requires effective feature selection to address high dimensionality and multicollinearity. This study proposes Iterative Partition Optimization (IPO), a framework integrating Model Population Analysis with Weighted Binary Matrix Sampling through segment-wise optimization: partitioning spectra into segments, isolating one active segment while freezing others, and using adaptive weighted sampling that learns from best-performing sub-models. Validation across four diverse NIR datasets (n=54-523 samples, 100-700 wavelengths) demonstrates IPO’s consistent performance improvement over conventional methods. For agricultural products (soy flour, wheat kernels), IPO achieved lower RMSECV while reducing wavelengths. In chemical analysis (diesel fuels, manure), the method maintained high prediction accuracy (RPD>3.0) using less than half the original variables. Notably in multi-component manure analysis, IPO improved predictions across seven chemical properties (N, NH4, P2O5, CaO, MgO, K2O, DM) while reducing spectral variables, consistently outperforming MCUVE in both accuracy and wavelength selection efficiency. These results establish IPO as an effective wavelength selection method for NIR spectroscopy, addressing multicollinearity while preserving spectral interpretation through optimized interval selection.

Author 1: Phuong Nguyen Thi Hoang

Author 2: Thinh Ngo Hung

Author 3: Tuong Nguyen Huy

Author 4: Hieu Nguyen Van

Keywords: Machine learning; feature extraction; near infrared spectroscopy; iterative partition optimization

Paper 83: A Secure Tracking System to Prevent Counterfeit Drugs in the Pharmaceutical Supply Chain

Abstract: Counterfeit drugs represent a serious global risk, accounting for about 10% of the medicines worldwide. This per-centage increases significantly in developing countries, reaching up to 30%, and has become a major cause of child mortality in these regions. Blockchain technology provides a good solution to address this critical issue through improving transparency and traceability of pharmaceuticals from manufacturers to end users. However, traditional blockchain applications face major challenges, particularly in terms of high costs and scalability especially when handling large volumes of data and transactions such as those found in the pharmaceutical supply chain. To address these limitations, researchers have introduced Layer 2 blockchain technologies, which operate on top of standard blockchain networks to improve scalability and reduce costs. This paper aims to design an efficient, secure, and scalable framework that can support traceability in pharmaceutical supply chains and contribute to reducing counterfeit drugs. This framework lever-ages Layer 2 blockchain solutions and combines Non-Fungible Tokens (NFTs) linked to Quick Response (QR) codes to provide unique identification for each drug. Furthermore, it incorporates the InterPlanetary File System (IPFS) for off-chain storage to reduce the amount of data stored directly on the blockchain. Additionally, an access control mechanism is incorporated to ensure the protection of sensitive data by limiting access based on predefined roles. This enhances privacy and reduces the risk of data leakage or misuse. In conclusion, this work provides the global scientific community with a secure and scalable framework that can be applied internationally to strengthen pharmaceutical supply chains and reduce the risks of counterfeit drugs.

Author 1: Huda Alrehaili

Author 2: Suhair Alshehri

Author 3: Rania M. Alhazmi

Keywords: Pharmaceutical supply chain; blockchain; access control; InterPlanetary File System (IPFS); Non-Fungible Token (NFT)

Paper 84: Integrating Large Language Models with Deep Reinforcement Learning for Portfolio Optimization

Abstract: This paper explores the application of Deep Reinforcement Learning (DRL) and Large Language Models (LLMs) to portfolio optimization, a critical financial task requiring strategies to balance risk and return in volatile markets. Traditional models often struggle with the complexity of financial markets, whereas Reinforcement Learning (RL) provides end-to-end frameworks for learning optimal, dynamic trading policies through sequential decision-making and trial-and-error interactions. The study examines key DRL algorithms, including Q-learning, Deep Q-Networks (DQN), Proximal Policy Optimization (PPO), and Twin-Delayed Deep Deterministic Policy Gradient (TD3), emphasizing their strengths in dynamic asset allocation. Crucial components of financial RL systems are discussed, such as state representations, reward function designs, its algorithms, and main approaches. Furthermore, the survey investigates how LLMs enhance decision-making by analyzing unstructured data (like news and social media) for sentiment and risk assessment, often integrating these insights to augment state representations or guide reward shaping within DRL frameworks.

Author 1: Renad Alsweed

Author 2: Mohammed Alsuhaibani

Keywords: Portfolio optimization; Reinforcement Learning (RL); algorithmic trading; Markov Decision Process (MDP); Large Language Models (LLM); deep learning

Paper 85: MOON Framework: An Emotionally Adaptive Voice Interaction Model for Older Adults

Abstract: As the world experiences a rapidly aging population, it has become a pressing design challenge to ensure that emerging digital technologies remain understandable, supportive, and meaningful for older adults. One of the most promising modalities for promoting inclusive interaction is the Voice User Interface (VUI), supported by artificial intelligence (AI) and speech recognition. However, most existing VUIs emphasize functional accuracy and overlook users’ emotional conditions, such as anxiety, confidence, or frustration, which significantly affect long-term engagement and adoption. To address this gap, this study introduces a structured and adaptive model for developing emotionally intelligent voice interfaces, namely MOON Framework (Model-Observation-Optimization-Nurture). This framework integrates demographic and linguistic profiling (Model), real-time emotional perception (Observation), adaptive vocal modulation (Optimization), and feedback-driven learning (Nurture) to create a closed-loop system capable of adjusting dynamically to user experience. Grounded in Affective Computing Theory, Socioemotional Selectivity Theory (SST), and Adaptive User Modeling, MOON Framework conceptualizes empathy as a measurable and computable dimension of VUIs. Unlike existing affective systems that rely primarily on visual or physical expressiveness, MOON demonstrates that vocal empathy, achieved through modulation of tone, cadence, and speech rate can foster comparable emotional attunement. Its cyclical design transforms emotion from a passive observation into an active driver of system adaptation. Focusing on familiarity and confidence specifically in older adults, MOON provides both a theoretical foundation and a practical framework for creating emotionally inclusive AI technologies that promote trust, engagement, and long-term well-being.

Author 1: Hasan Sagga

Author 2: Richard Stone

Keywords: Voice User Interfaces; emotional adaptivity; affective computing; older adults; human-computer interaction; empathy; accessibility; adaptive systems; MOON Framework

Paper 86: Informative Inputs Over Complex Features: Long Short-Term Memory Forecasting in the Saudi Stock Market

Abstract: This study investigates the effectiveness of deep learning, specifically Long Short-Term Memory (LSTM) networks, for forecasting stock closing prices in the Saudi Arabian market. Unlike prior research that focuses on narrow stock subsets or individual technical indicators, we present the first comprehensive evaluation across the top thirty companies by market capitalization on the Tadawul Exchange. We compare eight LSTM variants trained on different combinations of feature families, including technical indicators, calendar effects, and macroeconomic variables such as Brent oil prices and the Tadawul All Share Index (TASI). Despite the popularity of complex feature engineering, our results show that models using simpler macroeconomic inputs consistently outperform those based on technical indicators. In particular, combining the TASI with closing prices yielded the best results for 44.8% of stocks, improving median accuracy by 6.19% over the closing price-only baseline. Conversely, models incorporating extensive technical indicators or applying advanced feature selection techniques underperformed the baseline by 8 to 12%. These findings challenge the assumption that greater complexity leads to better performance in financial forecasting and highlight the value of focused, economically interpretable features in the Saudi market context.

Author 1: Munira AlBalla

Author 2: Arwa Alawajy

Author 3: Seetah Alsalamah

Author 4: Zahida Almuallem

Keywords: Deep learning; feature engineering; stock price forecasting; macroeconomic indicators; Brent crude oil; Saudi stock market; Long Short-Term Memory; Tadawul; Tadawul All Share Index (TASI)

Paper 87: Integrating Artificial Intelligence into Continuous Improvement for Automotive Manufacturing

Abstract: The integration of Artificial Intelligence (AI) and Machine Learning (ML) into Continuous Improvement (CI) frameworks is redefining the foundations of automotive manufacturing under the Industry 4.0 paradigm. Traditional method-ologies such as Kaizen, Lean Six Sigma, and Total Quality Management (TQM) have long provided structured approaches for quality enhancement, waste reduction, and process stability. However, the emergence of AI introduces new capabilities—advanced analytics, predictive modeling, and intelligent automation—that transform these static frameworks into dynamic, data-driven ecosystems. This study conducts a systematic literature review following the PRISMA protocol, covering publications from 2010 to 2024 across Scopus, Web of Science, and OpenAlex. After filtering and de-duplication, 13,080 documents were analyzed. Data were categorized by AI methodologies (computer vision, neural networks, deep learning), industrial use cases (quality inspection, predictive maintenance, process optimization, scheduling, and supply chain planning), and key performance metrics such as Overall Equipment Effectiveness (OEE), Mean Time Between Failures (MTBF), parts per million (ppm), lead time, and service level. The analysis reveals substantial and measurable performance improvements. AI-driven systems achieve an aver-age 15% gain in production efficiency, while computer vision enables automated defect detection, improving first-pass yield and reducing scrap. Predictive maintenance reduces unplanned downtime, increasing equipment availability and reliability. These benefits depend strongly on digital maturity and integration within enterprise systems—particularly Manufacturing Execution Systems (MES), Enterprise Resource Planning (ERP), and Product Lifecycle Management (PLM) which together ensure real-time data flow, process synchronization, and traceability across production operations. The primary barriers to adoption include data quality and governance issues, lack of workforce expertise, model explainability in safety-critical environments, and the complexity of integrating AI solutions into legacy systems. These factors hinder large-scale deployment despite proven technical advantages. This study proposes an applied framework for integrating AI within CI initiatives, aligned with the DMAIC (Define–Measure–Analyze–Improve–Control) cycle and the emerging Quality 4.0 architecture. It highlights managerial enablers such as data readiness, digital governance, and cross-functional collaboration, while identifying research gaps related to implementation costs, time-to-value, and long-term performance measurement. The findings demonstrate how AI transforms CI from reactive optimization to proactive, self-improving systems capable of sustaining excellence in modern automotive manufacturing.

Author 1: Sara OULED LAGHZAL

Author 2: EL OUADI Abdelmajid

Keywords: Industry 4.0; automotive manufacturing; artificial intelligence; machine learning; CNN; computer vision

Paper 88: Managerial Drivers and Performance Outcomes of AI Adoption in Automotive Manufacturing

Abstract: This article examines how artificial intelligence and machine learning reshape automotive manufacturing within Industry 4.0. Reported impacts include up to a 200 percent reduction in costs and a 400 percent gain in production efficiency, with controlled studies showing about a 15 percent improvement from process optimization. The largest early wins appear in quality management through computer vision and continuous inspection, followed by predictive maintenance that cuts unplanned downtime and stabilizes throughput. Supply chain and planning benefit from demand forecasting and inventory optimization that reduce bullwhip and working capital. Adoption barriers remain meaningful, including high initial investment, integration complexity, skills gaps, and trust and explainability requirements in regulated contexts. Effective programs use a common data and MLOps backbone, prioritize short cycle use cases, link model outputs to machine and recipe actions, and track value through OEE, ppm, MTBF, lead time, and service level. The discussion outlines practical steps to scale while noting evidence limitations and the need for standardized reporting on cost of ownership and time to value.

Author 1: Sara OULED LAGHZAL

Author 2: EL OUADI Abdelmajid

Keywords: Industry 4.0; automotive manufacturing; artificial intelligence; machine learning; quality 4.0; predictive maintenance

Paper 89: Mapping Artificial Intelligence Research in Automotive Manufacturing: A Bibliometric Study

Abstract: Artificial Intelligence (AI) is transforming the automotive industry by enabling smart manufacturing, optimizing supply chains, enhancing vehicle safety, and accelerating the shift toward autonomous mobility. This bibliometric study provides a systematic overview of the intellectual landscape of AI adoption in the automotive sector. Using data from Scopus, Web of Science, and OpenAlex, we analyze publication trends, influential authors, key institutions, thematic clusters, and international collaboration networks. Findings show a sharp rise in research output during the last decade, with major themes including predictive maintenance, computer vision for quality control, autonomous driving systems, supply chain optimization, and sustainable manufacturing. Emerging areas such as explainable AI, digital twins, and AI-enabled Industry 4.0 architectures are gaining increasing visibility. Collaboration analysis highlights strong contributions from Asia, Europe, and North America, with growing interdisciplinary networks bridging engineering, computer science, and management. This work not only maps the state of research but also identifies gaps and future directions for advancing AI adoption in the automotive industry. The study offers practical insights for researchers, industry practitioners, and policymakers aiming to harness AI for operational efficiency, competitiveness, and sustainable growth.

Author 1: Sara OULED LAGHZAL

Author 2: EL OUADI Abdelmajid

Keywords: Artificial Intelligence; automotive industry; Industry 4.0; autonomous vehicles; supply chain optimization; computer vision; predictive maintenance; digital twins; smart manufacturing; bibliometric analysis

Paper 90: AI in Web Development: A Comparative Study of Traditional Coding and LLM-Based Low-Code Platforms

Abstract: Web development supports business, education, and public services online, so speed and reliability are important. Low-code and no-code (LCNC) platforms aim to save time by using visual tools instead of writing all code. The impact of these platforms when combined with large language models (LLMs) has not been well studied. This paper compares a chatbot built in three coding stacks (Node.js, Python, Ruby) and one LCNC workflow in n8n that uses LLMs (Grok, Gemini, ChatGPT). The same tasks and prompts were used to test development time, speed, user ratings, and answer quality (precision, recall, F1). The study shows that LCNC with LLMs reduced build time by about 60 percent while keeping response speed close to hand-coded systems and reaching high answer quality (F1 up to 90 percent) with strong user approval. To clarify the main objective, the paper aims to evaluate whether LCNC+LLM integration offers a practical alternative to traditional coding approaches for intelligent web applications, particularly in terms of efficiency and maintainability. The challenge addressed is the limited empirical evidence comparing these two paradigms under identical conditions and using consistent performance metrics. Results are also interpreted relative to competing approaches in conventional development workflows, highlighting where LCNC tools match, exceed, or fall behind manual coding. Some areas, such as security and error handling, still require extra care and represent limitations of the present study. Overall, results show that LCNC with LLMs can be a useful way to build fast and reliable tools while lowering the development barrier for both developers and non-developers.

Author 1: Abiha Babar

Author 2: Nosheen Sabahat

Author 3: Arisha Babar

Author 4: Nosheen Qamar

Author 5: Marwan Abu-Zanona

Author 6: Asef Mohammad Ali Al Khateeb

Author 7: Bassam ElZaghmouri

Author 8: Saad Mamoun AbdelRahman Ahmed

Author 9: Lamia Hassan Rahamatalla

Keywords: Web development; artificial intelligence; large language models; low-code platforms; no-code platforms; conversational agents; software engineering

Paper 91: Towards a Tailored Cybersecurity Education Framework for Malaysia: A Systematic Literature Review

Abstract: Developing countries including Malaysia faces urgent challenges in cybersecurity education: preparing graduates who meet industry demands while addressing national cultural and regulatory contexts. Despite global advancements, no localized education framework currently aligns Malaysian higher education curricula with industry-required competencies. This systematic literature review (SLR) analyzed 65 academic and gray literature sources selected from an initial pool of 706 studies. The review employed thematic synthesis to examine the Malaysian cybersecurity programs incorporate technical competencies, policy literacy, and contextual relevance. Findings reveal four recurring gaps: limited integration of industry-aligned technical skills, insufficient adoption of hands-on pedagogies such as labs and gamification, underrecognition of professional certifications, and minimal incorporation of local policy and cultural considerations. These insights emphasize the necessity of a context-aware cyber-security education framework tailored for Malaysia. The study provides a conceptual foundation for designing an industry-driven curriculum model, supporting future research on cybersecurity competency development in higher education.

Author 1: Muhammad Asfand Yar

Author 2: Hock Guan Goh

Author 3: Kiran Adnan

Author 4: Ming Lee Gan

Author 5: Vasaki Ponnusamy

Keywords: Cybersecurity education; systematic literature review; Malaysia; industry competencies; higher education; curriculum model; hands-on learning

Paper 92: Bit Stability and Hash-Length Trade-Offs in Binary Face Templates

Abstract: Binary face templates are an appealing alternative to floating-point embeddings for face verification in resource-constrained environments because they enable constant-time Hamming matching with minimal storage and input/output (I/O). This paper studies the bit-level behavior of hashes obtained by principal component analysis followed by iterative quantization (PCA–ITQ) at L∈{32, 64, 128} derived from a frozen lightweight face encoder. Using subject-disjoint splits on the MORPH longitudinal dataset and an eyeglasses stress protocol on CelebA, the analysis quantifies 1) bit balance and entropy, 2) within-identity bit stability via per-bit flip rates, and 3) verification performance at low false-accept rates in Hamming space. On MORPH, 64-bit PCA–ITQ codes achieve an area under the receiver operating characteristic curve (AUC) of 0.9978 and a true positive rate (TPR) of 96.5% at a false positive rate (FPR) of 1%, compared to 99.1% at 128 bits, while halving the template length; 32-bit codes remain feasible but drop to 85.7% at the same operating point and are more sensitive to nuisance variation. Across both datasets, codes are near-balanced and mostly stable, yet a small minority of bit positions accounts for most flips under the eyeglasses attribute. In this regime, 64-bit hashes offer a favorable size–accuracy trade-off, whereas 128-bit hashes approximate float-embedding behavior and 32-bit hashes require redundancy or additional robustness mechanisms. All evaluations use fixed seeds and subject-disjoint splits; thresholds are selected on validation and held fixed on test to reflect deployment conditions.

Author 1: Abdelilah Ganmati

Author 2: Karim Afdel

Author 3: Lahcen Koutti

Keywords: Binary face templates; face verification; hamming distance; PCA-ITQ; bit stability; eyeglasses attribute; resource-constrained verification; MORPH; CelebA

Paper 93: A Quantized Deep Learning Model for Efficient Plant Leaf Disease Detection on Embedded Systems

Abstract: You Only Look Once (YOLO) object detection network has gained significant adoption in the field of plant leaf disease detection due to its strong detection capabilities. However, deploying YOLO models on resource-constrained devices remains challenging, as they require substantial computational power. The complexity and size of these models pose significant obstacles for edge platforms, which are often limited in processing and memory resources. To address these limitations and accelerate inference, we propose a quantized version of YOLOv5x, called Quant-YOLOv5x. This quantization reduces the size and complexity of the model, making it more suitable for edge deployment while maintaining competitive detection accuracy. The experiments were carried out using a self-generated dataset focused on detecting tomato and pepper leaf diseases. Our quantization method reduces the bitwidth of the entire YOLO network to 8 bits, resulting in only a 2.8% decrease in mean Average Precision (mAP), a 50% reduction in model size, and an increase of 4.7 FPS compared to the standard YOLOv5 model.

Author 1: Balkis Tej

Author 2: Soulef Bouaafia

Author 3: Mohamed Ali Hajjaji

Author 4: Abdellatif Mtibaa

Author 5: Mohamed Atri

Keywords: Object detection; plant disease; quantization technique; YOLOv5

Paper 94: AI-Assisted Schema Translation and Verified Migration: From Object-Relational Models to NoSQL Document-Oriented Stores

Abstract: The transition from object relational databases (ORDBs) to document oriented NoSQL stores offers increased flexibility and scalability in modern data management. However, existing migration processes remain largely manual and heuristic, hindering automation, formal verification, and adaptability to evolving workloads. This paper introduces a principled AI-assisted framework that unifies schema translation and data migration for end-to-end ORDB-to-NoSQL transformation. The framework operates through a four-stage pipeline comprising:(i) database metadata extraction, (ii) prediction of mapping strategies using a dual encoder that integrates a Graph Neural Network (GNN) for structural reasoning and a Transformer for semantic interpretation, (iii) automated data migration guided by the predicted mappings, and (iv) symbolic verification ensuring semantic and structural correctness. The verification stage enforces coverage, key fidelity, referential realizability, and type soundness constraints to guarantee loss-free transformation, while an optional workload-aware component refines verified mappings for query efficiency. Experimental evaluation on the RetailDB benchmark demonstrates that the proposed framework establishes a safe, explainable, and adaptable foundation for AI-assisted schema and data migration between object relational and document oriented models.

Author 1: Fouad Toufik

Author 2: Abd Allah Aouragh

Author 3: Abdelhak Khalil

Keywords: Graph Neural Networks (GNN); transformer models; neuro-symbolic verification; hybrid AI architectures; schema translation; data migration; object-relational databases; document-oriented databases

Paper 95: THMI-FS-Stack: A Hybrid Imputation and Feature Selection with Stacking Ensemble for Avian Influenza Outbreak Prediction

Abstract: Timely prediction of zoonotic disease outbreaks, particularly Highly Pathogenic Avian Influenza (HPAI), is critical for real-time epidemiological surveillance and pandemic pre-paredness. However, real-world avian surveillance datasets often suffer from missing values, high dimensionality, and inconsistent feature distributions, leading to unreliable predictions. This study proposes THMI-FS-Stack, a modular and interpretable machine learning pipeline that integrates hybrid data imputation, scalable feature selection, and ensemble classification for outbreak forecasting. The first stage, THMI-CB, employs a two-layer imputation framework combining statistical techniques (Mode, Hot Deck, KNN) and machine learning models (Bayesian Networks and CatBoost), achieving an F1-score of 0.91. The second stage, Hybrid-FS-ML, combines filter-based ranking (Mutual Information, Chi-Square, mRMR) with wrapper-based optimization using a Genetic Algorithm, achieving a 72% dimensionality reduction and an F1-score of 0.96. The final component is a stacking ensemble classifier that uses Random Forest and XGBoost as base learners and Logistic Regression as the meta-learner, yielding an F1-score of 0.92, accuracy of 0.93, and AUC-PR of 0.89. Evaluated on the Wild Bird HPAI dataset with 5-fold stratified cross-validation, THMI-FS-Stack consistently outperforms baseline models. Its robust architecture, low computational cost (runtime of 28–36s), and strong generalization ability make it highly suitable for noisy, incomplete epidemiological data in wildlife surveillance dashboards and early-warning systems.

Author 1: V. S. V. S. Murthy

Author 2: J. N. V. R. Swarup Kumar

Author 3: Srinivas Gorla

Keywords: Zoonotic disease outbreaks; avian influenza; pan-demic prediction; hybrid imputation; feature selection; stacking ensemble classifier; wild bird HPAI dataset; early-warning systems

Paper 96: Ensuring End-to-End Traceability and Sustainability in the FSC: A Modular Web3 Architecture Integrating Blockchain, IoT, and Machine Learning

Abstract: Traceability in food supply chains is crucial for ensuring safety, enabling effective quality control, and maintaining consumer trust. However, traditional paper-based or digital tracking systems often prove too slow and opaque during food safety incidents or investigations into fraud. To address these limitations, this paper presents a modular Web3 architecture that integrates Ethereum blockchain smart contracts, Internet of Things (IoT) sensors, and machine learning (ML) to achieve end-to-end traceability and sustainability in agrifood supply chains, and to support auditable, partially automated decision-making. The system design separates concerns into layers: an on-chain layer of Ethereum smart contracts for tamper-proof event logging and automated business logic, and an off-chain layer for secure storage of detailed sensor data and documents, linked by crypto-graphic hashes to ensure data provenance. Low-cost IoT sensors are deployed from farm to distributor, continuously monitoring environmental conditions (temperature, humidity, geolocation) and uploading signed, time-stamped summaries to the blockchain. In addition, ML models perform predictive quality control by estimating expected conditions, detecting anomalies, and scoring the conformity of product batches, which enables smart contracts to automatically trigger state transitions (acceptance or dispute escrow of shipments) based on real-time data. Using Ethereum smart contracts, a prototype that manages the life cycle of a specific food product was implemented, and two cases (conformant vs non-conformant shipments) were studied to demonstrate how cryptographically verifiable data and events make decisions transparent and trustworthy.

Author 1: Addou Kamal

Author 2: Mohammed Yassine El Ghoumari

Keywords: Food supply chain; traceability; blockchain; web3; smart contracts; IoT; machine learning; data integrity

Paper 97: A Hybrid Mutual-Information and Heatmap Driven Under-Sampling Algorithm for Imbalanced Binary Classification

Abstract: Class imbalance is a common problem that occurs in classification where one class has much more instances than the other class. Class imbalance is especially challenging in high-stakes fields like medical diagnosis, fraud detection and predictive maintenance among others. In cases of imbalanced class distribution, models perform well while predicting the majority class but are unable to predict the minority class, which is usually more important. This paper introduces the MI-Heat (Mutual-Information and Heatmap Driven Under-sampling), a hybrid algorithm which targets the binary classification problem. The algorithm is a data-level method that combines Mutual Information (MI) for identifying important features and K-Means clustering for identification of the most important majority class samples. In addition, a distance heatmap is used to project proximities among samples and cluster centers, guiding what majority instances to retain and what to discard. Together, Mutual Information, clustering, and heatmap preserve diversity and suppress noise in order to increase the ability of the model to represent both classes equally with clarity. Performance of the MI-Heat algorithm is tested on 23 benchmark datasets and the results are seen as an improvement in classification accuracy, minority class recall and model generalization. When compared to the traditional under-sampling approaches, MI-Heat performance is consistently better, which clearly demonstrates its dominance over dealing with the class imbalance issue.

Author 1: Sehar Gul

Author 2: Syahid Anuar

Author 3: Hazlifah Mohd Rusli

Author 4: Azri Azmi

Author 5: Sadaquat Ali Ruk

Keywords: Mutual information; heat-map visualization; clustering; under-sampling; data-level; hybrid

Paper 98: STROKECT-BENCH: Evaluating Convolutional and Transformer-Based Deep Models for Automated Stroke Diagnosis Using Brain CT Imaging

Abstract: Stroke detection from computed tomography (CT) images is an important research direction in computer vision. However, prior studies often use different preprocessing steps, model configurations, and evaluation protocols, making it difficult to compare results or assess architectural reliability. This paper presents an exploratory benchmark that evaluates representative convolutional neural networks (CNNs) and vision transformer (ViT) models under a unified experimental setting for binary stroke classification. STROKECT-BENCH is introduced as a standardized framework in which five CNNs and four transformer-based models are trained on the Brain Stroke CT Image dataset (1,551 normal and 950 stroke images) using identical preprocessing, augmentation, optimization parameters, and performance metrics. The results show that transformer models, particularly PVT-Small and Swin Transformer, achieve the highest accuracy and AUC, while EfficientNetB0 provides a strong balance between accuracy and computational efficiency. As an exploratory study, the findings aim to establish reliable baselines rather than clinical validation. STROKECT-BENCH offers a consistent evaluation reference for future work involving patient-level datasets, external validation, and multimodal stroke-analysis approaches.

Author 1: Raghda Essam Ali

Author 2: Reda Abdel-Wahab El-Khoribi

Author 3: Ehab Ezzat Hassanein

Author 4: Farid Ali Moussa

Keywords: Stroke detection; brain CT image; Convolutional neural networks; vision transformers; exploratory benchmark

Paper 99: Mobile Application Based on Clean Architecture for Sustainable Tourism Recommendation Using BERT

Abstract: The development of intelligent systems requires robust architectures and design patterns that ensure scalability, maintainability, and separation of concerns. This paper presents the design and implementation of a mobile application for recommending sustainable tourist destinations in Peru, employing a Clean Architecture approach in the backend and the MVC pattern in the frontend. The solution integrates a content-based recommendation engine using semantic embeddings generated through BERT, connected to open data sources from MINCE-TUR, SafeTravels, and OpenStreetMap. The proposed system was evaluated using three simulated user profiles, achieving an average similarity score of 0.83 and a more balanced distribution between traditional and emerging destinations. From a software engineering perspective, the system demonstrates decoupling between layers, modularity, interoperability with external APIs, and scalability potential. This architecture serves as a practical case study of applying modern design patterns to intelligent systems with an impact on sustainable tourism.

Author 1: Cristhian Paolo Atuncar Yataco

Author 2: Marco Antonio Lopez Salinas

Author 3: Jos´e Alfredo Herrera Quispe

Keywords: Software engineering; clean architecture; MVC; BERT; recommendation systems; mobile applications; FastAPI; Kotlin; sustainability

Paper 100: A String-Similarity-Oriented Item Swapping Algorithm for Protecting Sensitive Frequent Itemsets in Transaction Databases

Abstract: Frequent itemset mining is a widely adopted data mining technique. The application of the technique can be found in transaction database analysis, such as exploring a set of purchased items. Presently, the growing concern of privacy protection and security issues in society leads the business parties to be more careful in handling their database, since various information can be extracted from the database including the sensitive information. Therefore, database owner should take measure to minimize sensitive information leak during the data mining process. A hiding sensitive frequent itemset algorithm can be adopted to achieve that. However, it is remain a challenge to design a data hiding algorithm that not only successfully hiding frequent sensitive itemset but also minimize the side effects such as item loss, the appearance of artificial frequent itemset and misses cost. In this paper, a method namely D-LSwap that based on item swapping technique is proposed to cope the issue while minimizing those side effects. Initially D-LSwap inspect each transaction in the database to determine whether a transaction is sensitive. Following that, it select a sensitive transaction from the previous process and create a pair of transactions from them. The pair is formed by incorporating Damerau-levensthein string similarity. The next step is selecting items from this pair for the swapping process. Experiment results indicate that the proposed method outperforms several existing algorithms by increasing data utility up to 10%, while minimizing the number of item loss more than 10 times lower than that of the baseline methods.

Author 1: Fatah Yasin Al Irsyadi

Author 2: Dedi Gunawan

Author 3: Diah Priyawati

Author 4: Wardah Yuspin

Keywords: Data hiding; sensitive frequent itemset; frequent itemset; data mining; item swapping technique; D-Lswap

Paper 101: A Novel Access Control Model with the Skew Tent Map for Decision Making (STM-ABAC)

Abstract: The proliferation of cloud computing exposes sensitive data to the risk of unauthorized access, as traditional access control mechanisms are often inadequate for this dynamic environment. To address these shortcomings, this article proposes a novel access control scheme, named STM-ABAC, which is based on the Skew Tent Map (STM). This scheme is specifically designed to overcome the inherent limitations of traditional Attribute-Based Access Control (ABAC) and Attribute-Based Encryption (ABE) schemes when deployed in dynamic cloud environments. The methodology involves constructing a multi-authority ABAC model, generating verifiable attribute tokens using chaotic sequences, applying LSSS-based policy encryption, and evaluating performance through rigorous formal analysis and experimental benchmarking. The results demonstrate that STM-ABAC reduces the computational overhead during decryption by up to 60% and maintains lower initialization and key-generation costs compared to existing CP-ABE and MA-ABE schemes. Furthermore, security proofs confirm strong resistance to chosen-attribute and chosen-nonce attacks.

Author 1: Omessead BenMbarak

Author 2: Anis Naanaa

Author 3: Sadok ElAsmi

Keywords: Cloud computing; skew tent map; ABAC; ABE; attribute token

Paper 102: A Novel Approach for Urban Traffic Congestion Prediction

Abstract: Traffic congestion is a global problem in urban areas that creates longer travel times, increased fuel consumption, and elevated levels of pollution. Traffic congestion occurs because of the exponential growth of vehicles along with a finite number of roadways and the inability to manage traffic effectively. This paper studies the question: How well can traffic type factors be used as a predictor for determining the severity of traffic con-gestion? To answer this question, we present a new methodology to perform clustering and classification based on various types of traffic indicators. In addition, traffic indicators (such as size of roadway, speed of vehicles, number of vehicles, and level of traffic flow) are categorized by using two distinct classifications: homogeneous and heterogeneous. Using these categories, we then apply a modified version of the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm to do clustering of traffic indicators. The resultant label from the clustering process is then used to develop a prediction model that will provide information regarding the level of traffic congestion along a selected roadway. Results from our experiments were conducted using an actual dataset and demonstrate that our proposed method produced an accuracy rate of 93% with 92%precision and recall, and therefore, outperforming other current methodologies used for predicting traffic congestion. Overall, these findings indicate that incorporating an analysis of traffic type factors into the clustering and classification methodology can result in more accurate predictions of traffic congestion.

Author 1: Chaimae Kanzouai

Author 2: Abderrahim Zannou

Author 3: Soukaina BOUAROUROU

Author 4: El Habib Nfaoui

Author 5: Abdelhak Boulaalam

Keywords: Traffic congestion; traffic management; traffic factors; congestion level; DBSCAN; GCN

Paper 103: Lightweight Cryptography for Energy-Conscious Authentication in IoT Systems

Abstract: The proliferation of IoT devices has introduced new challenges in ensuring secure communication while maintaining computational and energy efficiency. Traditional cryptographic mechanisms such as AES-128, although robust, often fail to meet the stringent constraints of resource-limited IoT environments. This research investigates lightweight cryptographic algorithms and energy-conscious authentication protocols as viable alternatives for enhancing security in IoT systems. A comprehensive performance evaluation is conducted through the simulated implementation of three widely used algorithms—AES-128, SPECK, and PRESENT—measuring encryption time, decryption time, memory consumption, and ciphertext size. The results reveal critical trade-offs between security strength and resource usage, highlighting SPECK’s balanced performance and PRESENT’s ultra-low resource footprint. A Multi-Metric Suitability Analysis framework is introduced to assess overall applicability across different IoT use cases. The study provides evidence-based insights to guide algorithm selection for secure yet efficient IoT deployments. Future directions include hardware-based testing, integration into IoT protocols, and exploration of post-quantum lightweight cryptographic approaches. The findings contribute to the development of practical, scalable, and energy-efficient security architectures tailored for the next generation of IoT systems.

Author 1: Ashwag Alotaibi

Author 2: Huda Aldawghan

Author 3: Mounir Frikha

Keywords: Lightweight cryptography; IoT security; energy efficiency; lightweight authentication; SPECK; PRESENT; AES- 128; cryptographic algorithms; resource-constrained devices; Multi-Metric Suitability Analysis

Paper 104: Hybrid Reinforcement Learning-Based Hyper-Parameter Optimization with Yolov8 Indoor Fire Recognition

Abstract: This study presents an indoor vision-based fire detection system that integrates a YOLOv8n object detection model with a Reinforcement Learning-based Optimization Algorithm (ROA) for hyperparameter tuning. The research investigates three key aspects: 1) the effectiveness of ROA in improving model performance, 2) the optimal smart camera resolution and placement for indoor fire detection, and 3) the implementation of a real-time dual-channel user notification system. The BantaySunog model iteratively adjusted a few hyperparameters using the Reinforcement Learning-based Optimization Algorithm (ROA) [Talaat & Gamel, 2023]. An episodic framework was used for training, with 15 episodes of 20 epochs each, for a maximum of 300 epochs. Each episode's top weights were carried over to the following one. To balance exploration and exploitation, ROA employed an epsilon-greedy policy with an epsilon value that decreased from 0.9 to 0.2. Experimental results show that while ROA reduced training time and yielded a more conservative prediction strategy, it did not consistently outperform the baseline YOLOv8n in terms of detection metrics such as recall and mAP50-95. Camera deployment tests identified that positioning cameras away from direct light sources significantly improved detection success, with both elevation and resolution contributing to overall system performance. Finally, a dual-channel alert mechanism combining Firebase Cloud Messaging (FCM) and Telerivet SMS API enabled the timely delivery of fire alerts, aligning with real-world standards. The findings contribute to the development of reliable and accessible fire detection systems, especially for densely populated residential areas with limited infrastructure.

Author 1: Patrick D. Cerna

Author 2: Harvey C. Quijada

Author 3: Michael Joseph E. Ortaliz

Author 4: Kenneth Charles H. Saluna

Keywords: Fire detection; YOLOv8; reinforcement learning; hyperparameter optimization; convolutional neural network; indoor surveillance; real-time alert system

Paper 105: Adaptive Open Cyber Intelligence for SOAR: Reduced False Positives in Low-Resource Environments

Abstract: The increasing use of resource-constrained cyber-physical devices emphasizes the need for effective and flexible methods in the deployment of threat intelligence. The Open Cyber Intelligence Framework (OCIF), an architecture that applies Security Information and Event Management (SIEM) and Security Orchestration, Automation, and Response (SOAR) capabilities to resource-constrained environments, is presented in this study. The OCIF uses lightweight machine learning models in an adaptive way to process cyber threat intelligence (CTI) with greater precision and effectiveness. By using Wazuh to monitor the behavior of machines and OpenSearch for modeling the results of the analysis, the OCIF can reduce false positives by up to 6% in real-world implementations. The model ensures sufficient threat mitigation without taxing the system by striking a balance between anomaly detection, context, and decreased communication overhead. Because of its open-source propagation and modular form factor, OCIF promotes innovation and makes it possible for CTI to be built and used in restricted resources with optimal detection and operational efficiency.

Author 1: Shunmugam U

Author 2: Rajesh D

Keywords: Open Cyber Intelligence Framework; SOAR; SIEM; Cyber Threat Intelligence; false positive reduction; threat mitigation; anomaly detection