IJACSA Volume 17 Issue 1 - thesai.org

IJACSA Volume 17 Issue 1

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: A Multi-Model Adaptive Q-Learning Framework for Robust Portfolio Management in Stochastic Markets

Abstract: This study presents TAQLA, a new Tabular Adaptive Q-Learning Agent for portfolio management in stochastic financial markets. TAQLA rests on a multi-model reinforcement learning (RL) architecture that integrates parameter-adaptive Q-Learning mechanisms into softmax-based exploration to reconcile short-term profit maximization with long-term capital preservation. The method is contrasted with vanilla Q-Learning, SARSA, and a random trading policy using simulated equity market data. Empirical analysis shows that TAQLA performs better on profitability, risk-adjusted performance, and drawdown minimization, with a last portfolio value of $1687.45 (+68.74% of initial capital), a Sharpe ratio of 1.41, and a maximum drawdown of just 12.8%. Q-Learning and SARSA, on the other hand, yield Sharpe ratios below 1.0 and drawdowns exceeding 18%. Parameter sensitivity analysis across β (softmax temperature), α (learning rate), and γ (discount factor) reveals that aggressive exploration (β ≈ 1.0–1.5) and reasonable discounting (γ ≈ 0.4–0.6) generate the most aggressive and robust outcomes. Such outcomes place TAQLA as a robust RL-based adaptive portfolio control method under uncertainty, with improved capital appreciation and robustness to adverse market conditions.

Author 1: Sharmin Sultana

Author 2: Md Borhan Uddin

Author 3: Masuma Akter Semi

Author 4: Shahanaj Akther

Author 5: Urmi Chakraborty

Author 6: Khandakar Rabbi Ahmed

Keywords: Reinforcement learning; Q-Learning; tabular reinforcement learning; portfolio management; dynamic asset allocation

Paper 2: AI-Driven Anomaly Prediction in Encrypted Network Traffic

Abstract: The rapid growth of computer networks has increased demand for more sophisticated tools for network traffic analysis and monitoring. The increasing reliance on networks has amplified the need for robust security and intrusion detection mechanisms. Numerous studies have sought to develop efficient methods for fast and accurate intrusion detection, each addressing the challenge from different perspectives. A common limitation among these approaches is their reliance on expert-engineered features extracted from network traffic. This dependency makes them less adaptable to emerging attack techniques and changes in normal traffic patterns, often resulting in suboptimal performance. In this study, we propose a method leveraging recent advancements in artificial neural networks and deep learning, specifically using recurrent neural networks (RNNs), for network traffic analysis and intrusion detection. The key advantage of this approach is its ability to autonomously extract features from network traffic without human intervention. Trained on the ISCX IDS 2012 dataset, the proposed model achieved an accuracy of 0.99 in distinguishing between malicious and normal traffic.

Author 1: Sina Ahmadi

Keywords: Machine learning; deep learning; recurrent neural networks; intrusion detection

Paper 3: An Intelligent Knowledge-Based Chatbot to Mitigate Travel Anxiety

Abstract: With the emergence of intelligent chatbots, AI-driven conversational agents are increasingly being used to help tourists manage travel challenges and obtain effective solutions. Travel anxiety constitutes a significant impediment to tourism, substantially influencing travelers' future intentions. Given its multifaceted nature, spanning from subjective experiences to complex logistical arrangements, this study developed a fully functional tourism chatbot system using a tourist-centered design method to provide targeted guidance for travel anxiety mitigation. The knowledge-based chatbot implemented user-centered evaluation methods by recruiting seven participants who were randomly assigned to scenarios across six major global travel regions. Results from the participants’ short-answer responses and Likert-scale usability ratings indicated that this knowledge-based system delivers highly informative, context-aware, and expert-level recommendations through multifaceted strategy implementation. The findings suggest that such AI-driven interventions are effective in addressing specific travel challenges, with further implications for user-centered design discussed herein.

Author 1: Jieyu Wang

Author 2: Hungchih Yu

Author 3: Dingfang Kang

Keywords: Tourist-centered design; user-centered evaluation; knowledge base; context-aware chatbot; travel anxiety

Paper 4: Towards Robust Intrusion Detection: Exploring Feature Selection, Balancing Strategies, and Deep Learning for Minority Class Optimization

Abstract: The increasing connectivity of systems and the rapid growth of the Internet have intensified cybersecurity threats. It has been demonstrated that conventional signature-based intrusion detection methods are deficient, especially against Zero-Day attacks. An alternative approach involves the deployment of Intrusion Detection Systems (IDS) that are based on deep learning algorithms. However, these systems face a significant challenge in detecting minority classes of attacks, such as Remote-to-Local (R2L) and User-to-Root (U2R) attacks, which, although rare, are of critical importance. Misclassifying these attacks is costly. Therefore, the reduction of false negatives is achieved by coupling feature selection techniques (Chi square, correlation, information Gain, Extreme Gradient Boosting (XGBoost), Autoencoder), oversampling methods (Synthetic Minority Oversampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN)) and deep learning models (Deep Neural Network (DNN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) and hybrid model CNN LSTM). The present study uses the NSL-KDD dataset, with a particular focus on the minority classes R2L, which represents 2.61% of the dataset, and U2R, representing 0.08% of the dataset. The findings indicate that data balancing is paramount. ADASYN facilitates 100% U2R detection, while SMOTE enhances R2L accuracy to above 95%. The application of correlation and autoencoder feature selection techniques proved to be the most effective. The effectiveness of CNN models in addressing U2R classification tasks has been extensively demonstrated, while the use of DNN or CNN-LSTM models has been shown to yield optimal results for R2L tasks. DNN remains the most stable model overall. For the two minority classes, the most effective pipelines are Correlation + SMOTE + DNN, achieving 93.84 % recall for U2R and 99.88 % for R2L, and Autoencoder + SMOTE + CNN-LSTM, achieving 89.66 % recall for R2L and 99.68 % for U2R.

Author 1: Khalid LABHALLA

Author 2: Amal BATTOU

Keywords: Network intrusion detection system; imbalanced data; minority class detection; deep learning; feature selection; balancing techniques

Paper 5: DriveRight: An Embedded AI-Based Multi-Hazard Detection and Alert System for Safe and Sustainable Driving

Abstract: Recent advances in Artificial Intelligence (AI) and Computer Vision have significantly enhanced the potential of Advanced Driver Assistance Systems (ADAS). However, existing solutions remain limited by high computational cost, single-function design, and dependence on expensive sensors such as radar and LiDAR. This study presents DriveRight, an embedded AI-based driver-assistance system that integrates multi-scenario hazard detection and real-time object detection and alerting using a single low-cost vision sensor on a Raspberry Pi platform. The system leverages a simulation-to-deployment pipeline, combining CARLA-based synthetic training environments with TensorFlow deep learning models, including SSD Inception v2, MobileNet-SSD, and Faster R-CNN. Experimental results show that Faster R-CNN achieved 92.1% detection accuracy for vehicles and 90.3% for traffic signs, while MobileNet-SSD achieved real-time performance at 14.6 frames per second (FPS) with minimal latency of 2.8 seconds on embedded hardware. Field tests validated the system’s ability to accurately detect and classify stop signs, vehicles, and lane deviations under varying lighting and motion conditions, triggering timely alerts to the driver. The prototype demonstrates a cost-effective and energy-efficient AI solution (< 12 W) for intelligent transportation systems. The findings establish the feasibility of deploying IoT-based ADAS and deep learning–driven driver-assistance technologies in low-cost, sustainable embedded platforms, bridging the gap between research-grade ADAS and practical real-world deployment.

Author 1: Jamil Abedalrahim Jamil Alsayaydeh

Author 2: Rex Bacarra

Author 3: Ahamed Fayeez Bin Tuani Ibrahim

Author 4: Mazen Farid

Author 5: Aqeel Al-Hilali

Author 6: Safarudin Gazali Herawan

Keywords: Embedded AI; computer vision; intelligent transportation; IoT-based ADAS; deep learning; real-time object detection; Raspberry Pi

Paper 6: Towards an AI-Powered Cyber Resilience Model: A Systematic Evaluation of Frameworks Against Emerging Threats

Abstract: This study presents a Systematic Literature Review of cyber resilience frameworks against emerging threats, published between 2010 and 2025. While numerous frameworks exist, their ability to anticipate, withstand, and evolve in the face of sophisticated attacks remains uncertain. The study maps frameworks across nine resilience goals, namely Identify, Protect, Detect, Respond, Recover, Govern, Anticipate, Withstand, and Evolve, creating a goal-wise evidence matrix and quantification. Using the PRISMA methodology, 11,027 publications were identified, of which 55 studies met the inclusion criteria for critical analysis. The results indicate that most frameworks accentuate Protect and Detect functions at 87.72 per cent, whereas Govern at 17.54 per cent, Withstand at 28.07 per cent, and Evolve at 24.56 per cent remain under-represented. Only 45.61 per cent of frameworks explicitly address emerging threats such as Artificial Intelligence-driven or Internet of Things-based attacks. Strengths observed include situational awareness, Artificial Intelligence and Machine Learning integration, dynamic defence mechanism, Blockchain, and adoption of Zero Trust principles. The key weaknesses lie in the undervalued cyber resilience goals, namely Govern, Withstand, and Evolve, low empirical validation, and a narrow scope in addressing emerging threats, which highlight gaps that limit resilience against sophisticated attacks. Based on these findings, an evidence-informed Artificial Intelligence-powered cyber resilience model is proposed that privileges adaptability and future proofing. This review highlights the urgent need for cyber resilience frameworks to expand beyond reactive measures and to embed forward-looking resilience capabilities.

Author 1: Chhaya Jahajeeah-Suntoo

Author 2: Sheeba Armoogum

Keywords: Cyber resilience; cybersecurity framework; Artificial Intelligence; emerging threats; Zero Trust; systematic literature review

Paper 7: Dual Cognitive Pathway Architecture for Robust and Dialect-Aware English Reading Comprehension

Abstract: Reading comprehension models frequently struggle to accommodate linguistic diversity, especially dialectal variations within the English language that disrupt semantic alignment and fairness. In order to overcome these drawbacks, this study proposes the Dual Cognitive Pathway-Based Dialect-Aware Cognitive Twin Framework (NeuroTwin-DialectaLearn) as a new framework that combines the principles of cognitive science, sociolinguistic expertise, and adaptive learning methods. The framework also has two parallel understanding routes, one is a Lexico-Semantic Pathway that processes normal English, and the other is a Dialectal-Semantic Pathway that is involved in normalizing the dialect and aligning the semantics. Such pathways interact with each other by a process of Adaptive attention fusion in a Cognitive Twin model, which instantiates important cognitive processes, including lexical processing, syntactic parsing, semantic integration, inductive reasoning, and answer generation. It uses Python and PyTorch to implement the system and is tested on the English Classroom QA Dataset with the addition of synthetic dialectal variants to enhance the system. The accuracy of comprehension is 98.1 per cent, the average response time is 14.3 seconds, and the rate of learners’ improvement is 18.9 per cent, which is much higher than the baseline QA systems and improved BERT models. The framework has shown consistent performance in the context of dialects; the number of vocabulary-related mistakes is lower, and the consistency of inference is higher, proving to be an effective tool in dialect-conscious and cognitively based reading comprehension. In general, the present research provides a linguistically encompassing, versatile, and understandable next-generation smart educational system.

Author 1: Jillellamoodi Naga Madhuri

Author 2: Tonmoyee Doley

Author 3: Purnachandra Rao Alapati

Author 4: Vijaya Kumar P

Author 5: Linginedi Ushasree

Author 6: Marvin D. Mayormente

Author 7: Aseel Smerat

Keywords: Cognitive Twin; Dialectal Transfer; Adaptive Tasking; English reading comprehension; personalized learning

Paper 8: Proposed Technological Solution to Predict the Need for Health Professionals in Health Centers Using Random Forest

Abstract: The objective of this research is to develop a technological solution based on the Random Forest algorithm to predict healthcare workforce requirements in public healthcare centers in Peru, addressing staff shortages and unequal workforce distribution. A national dataset from the Peruvian Ministry of Health (MINSA) covering the period 2017–2024, segmented by levels of care (I, II, and III), was used to capture the operational differences within the healthcare system. The model, validated using an 80/20 split, achieved outstanding performance, with coefficients of determination (R²) exceeding 0.99 and minimal percentage errors (MAPE) across all levels of care. The main contribution of this work lies in converting estimated healthcare attendances into an operational metric of “required healthcare professionals”, integrated into a web-based architecture built on React, Flask, and PostgreSQL. The findings identify medical specialty and year as the most influential predictive variables. It is concluded that the proposed tool is robust for optimizing strategic healthcare workforce planning, enabling a more equitable and data-driven allocation of medical specialists.

Author 1: Fiorella Patricia Mirano Surquislla

Author 2: Gianfranco Henry Ore Paredes

Author 3: Aguilar-Alonso Igor

Keywords: Technological solution; Random Forest; healthcare sector; healthcare professional prediction; human resource management

Paper 9: A Readability-Driven Prompting Framework for Accurate Grade-Specific EFL Narrative Creation

Abstract: The integration of Artificial Intelligence (AI) into English as a Foreign Language (EFL) education offers new opportunities for developing adaptive and engaging learning materials. Narrative-based content is central to improving reading comprehension, vocabulary acquisition, and learner motivation. However, maintaining grade-appropriate readability in AI-generated narratives remains a major challenge. This study presents Readability-Driven Prompting (RDP), a novel technique designed to enhance the accuracy and efficiency of large language models in generating grade-level narratives. Using GPT-4o-mini, three prompting strategies—CEFR Keyword-Constrained Prompting (CKCP), Instruction-Based Prompting (IBP), and the proposed RDP—were applied to produce narratives for 7th-grade (A1–A2 CEFR) and 10th-grade (B1–B2 CEFR) learners. The outputs were evaluated using Flesch Reading Ease (FRE), Dale–Chall (DC) readability metrics, lexical analysis, and human assessments. Experimental results indicate that the RDP approach achieves higher alignment with target readability levels and improved lexical appropriateness compared to baseline methods, demonstrating a scalable and effective strategy for generating educational narratives, particularly for beginner-level learners.

Author 1: Ronald William Marbun

Author 2: Makoto Shishido

Keywords: Artificial Intelligence (AI); English as a Foreign Language (EFL); large language models (LLMs); readability metrics; narrative generation; prompt engineering; educational technology

Paper 10: HELM-BRCA: Hybrid Embedding and Learning Model for BRCA Methylation Classification

Abstract: Breast cancer remains a highly heterogeneous disease for which it demands advanced computational techniques that can reveal significant biological patterns in high-dimensional epigenomic data. DNA methylation profiles generated by the Illumina HumanMethylation450 platform yield rich, clinically relevant signals but introduce significant analytical challenges due to their high dimensionality, sparsity, and nonlinear structure. This work presents a novel memory-efficient hybrid learning architecture that combines Truncated Singular Value Decomposition (SVD), a deep Autoencoder, and a multi-model ensemble classifier for boosting subtype classification performance using TCGA-BRCA methylation data. In order to circumvent memory limits and prevent system crashes, a probe-subset extraction strategy combined with variance-based feature selection was employed to ensure fast and safe data loading from the Xena repository. While the autoencoder extracts compact nonlinear manifold representations, SVD captures the global linear variance structure. Further, the fused latent space is modelled by an ensemble including Random Forest, XGBoost, and a lightweight Keras neural classifier that allows the system to exploit different decision limits and achieve robust generalization. The experimental investigation across several architectures demonstrates high predictive performance with ROC-AUC scores exceeding 0.99 and accuracies higher than 0.96 for Basic CNN and MLP models. Furthermore, the proposed hybrid ensemble improves stability and precision by outperforming traditional baselines and confirming the complementary nature of spectral and deep feature extraction. This study is suitable for large-scale biomedical data analytics scenarios. In conclusion, this work provides an efficient hybrid machine learning framework for breast cancer methylation study by offering a strong platform for improved prognostic modelling and development of epigenetic biomarkers.

Author 1: Hemalatha D

Author 2: N Gomathi

Keywords: Breast cancer classification; DNA methylation; TCGA-BRCA; Truncated SVD; autoencoder; ensemble learning; deep learning; epigenomic biomarkers; hybrid model; machine learning pipeline; high-dimensional data

Paper 11: Strengthening Indonesia’s Unmanned Aerial Vehicle Manufacturing Industry: A Technology-Focused Strategic Analysis

Abstract: The development of national Unmanned Aerial Vehicle (UAV) technology represents a strategic imperative that requires immediate implementation. This study provides strategic recommendations to strengthen Indonesia’s UAV industry by employing SWOT analysis, the Analytical Hierarchy Process (AHP), and the Quantitative Strategic Planning Matrix (QSPM). Sixteen key internal and external factors were identified, with SWOT mapping situating UAVs in Quadrant I (Aggressive Strategy) at coordinates +1.24 and +0.60. AHP prioritization indicates that the strengths–opportunities (S–O) strategy (0.348) is of highest importance, emphasizing infrastructure enhancement and the adoption of advanced technologies. IFAS–EFAS integration confirms Wulung UAV’s aggressive growth position, while internal strengths account for 37.1% of overall strategic influence. QSPM analysis further validates the S–O strategy as optimal, with the highest internal (4.88) and external competitive (4.63) impact scores. Implementation of this strategy necessitates immediate action focused on manufacturing infrastructure enhancement, technological adoption, development of technical human capital, organizational capability strengthening, establishment of a domestic supply chain and supporting industries, and enforcement of robust industrial governance.

Author 1: Satrio Utomo

Author 2: Gani Soehadi

Author 3: Sarjono Sarjono

Author 4: Yanuar Iman Dwiananto

Author 5: Adi Akhmadi Pamungkas

Author 6: Ellia Kristiningrum

Author 7: Budi Setiadi Sadikin

Author 8: Hardono Hardono

Author 9: Rizki Arizal Purnama

Author 10: Helen Fifianny

Keywords: Unmanned Aerial Vehicle; technology adoption; strategic development; SWOT; AHP; QSPM

Paper 12: AI Mathematical: Solving Math Challenges Using Artificial Intelligence Models

Abstract: Artificial Intelligence (AI) has emerged as a transformative tool for solving mathematical challenges across diverse domains, ranging from algebra and geometry to calculus and number theory. This study investigates the role of AI in mathematics by analyzing three representative platforms—MathGPT.org, Math-GPT.ai, and StudyX.ai—and by proposing ten Python-based problem-solving models tailored to Olympiad-style problems. The methodology integrates rule-based reasoning, brute-force search, and heuristic strategies, while benchmarking is inspired by the AI Math Olympiad (AIMO) Progress Award competition on Kaggle. A comparative evaluation was conducted to assess accuracy, reasoning depth, and computational efficiency. Results show that AI solvers can provide step-by-step solutions, interactive visualizations, and adaptive learning support, but their performance varies depending on problem type and strategy. This study highlights both the potential and limitations of AI in mathematics education and research, emphasizing the need for automated model selection (AutoML) and formal benchmarking to strengthen credibility. The findings demonstrate that AI can simultaneously promote automated problem-solving and enhance personalized STEM learning.

Author 1: Trinh Quang Minh

Author 2: Ngo Thi Lan

Author 3: Bui Xuan Tung

Author 4: Phan Thanh Tuyen

Keywords: AI math solvers; Artificial Intelligence; STEM education; MathGPT.org; Math-GPT.ai; StudyX.ai; Python models; Olympiad problems; automated reasoning; Kaggle AIMO Progress Prize

Paper 13: A Comprehensive Forensic Framework for Unmanned Aerial Vehicle Investigations: Empirical Validation with the DJI Mavic 3 Classic

Abstract: The rapid proliferation of unmanned aerial vehicles has introduced significant challenges for digital forensics, particularly due to their increasing involvement in criminal, surveillance, and security-related incidents. The heterogeneous hardware architectures, proprietary data formats, encryption mechanisms, and volatile storage characteristics of modern drones complicate reliable evidence recovery and analysis. This study proposes a comprehensive forensic framework for unmanned aerial vehicle investigations, empirically validated using the DJI Mavic 3 Classic. The proposed methodology integrates a conceptual forensic model with practical investigation procedures, including multi-source data acquisition, metadata analysis, anomaly detection, and digital twin-based reconstruction to support event correlation and timeline reconstruction. Four representative case studies: flight log recovery, firmware modification detection, metadata-driven espionage analysis, and reconstruction of deleted media are conducted to evaluate the framework’s effectiveness. Experimental results demonstrate evidence recovery rates of up to 92%, timeline reconstruction accuracy of 95%, and anti-forensic activity detection rates of 100%. The framework explicitly addresses challenges associated with proprietary formats, encryption, and data volatility in drone ecosystems. The proposed approach provides actionable guidance for drone forensics practitioners, researchers, and policymakers, contributing toward standardized and reliable forensic investigation processes for contemporary unmanned aerial vehicle platforms.

Author 1: Nidhiba Parmar

Author 2: Naveen Kumar Chaudhary

Keywords: Drone forensics; unmanned aerial vehicle; DJI Mavic 3 Classic; digital forensics; metadata analysis; digital twin; anomaly detection; evidence recovery

Paper 14: Correlation Characteristics-Based Channel Estimation Method for GFDM Systems

Abstract: Generalized Frequency Division Multiplexing (GFDM) has broad application prospects due to its flexible subcarrier structure and low out-of-band leakage. Traditional channel estimation methods for GFDM systems rely on inserting a large number of pilot sequences, which reduces the data transmission rate. To address this problem, a channel estimation method for GFDM systems based on subcarrier correlation is proposed. First, according to the time–frequency characteristics of the prototype filter in the GFDM system, a pilot sequence with a two-dimensional time–frequency block structure (CTFP) is designed. This sequence is adjusted based on the parameters of the prototype filter. Then, the correlation among subcarriers is utilized for channel estimation, which effectively reduces the pilot overhead and improves the data transmission rate and interference resistance of the system. Simulation results show that under the same total time slot overhead, the mean square error and bit error rate performance of the proposed correlation-based methods are similar to those of existing methods, while the data transmission rate is improved by 14.97% compared with conventional methods.

Author 1: Xiaotian Li

Author 2: Xiaoqing Yan

Author 3: Zitian Zhao

Author 4: Jiameng Pei

Keywords: GFDM; channel estimation; correlation; pilot

Paper 15: ForenVoice-Secure: Robust and Privacy-Aware Audio Data Mining for Forensic Speaker Identification

Abstract: Speech is now routine evidence in criminal investigations, but forensic audio rarely matches the clean assumptions of standard speaker recognition. Clips are short, noisy, codec-compressed, and channel-mismatched, and they are increasingly exposed to replay and synthetic speech manipulation. Therefore, the cast criminal voice identification is forensic audio data mining, aiming to extract a stable identity structure from heterogeneous and potentially adversarial evidence, while respecting operational and privacy constraints. In this study, a novel ForenVoice-Secure system is proposed, a unified pipeline that combines robust representation learning, spoof-aware decisioning, and privacy-preserving training. Audio is mapped to log-Mel spectrograms and encoded with a CNN, while an LSTM aggregates temporal identity cues from irregular utterances. Robustness is improved through multi-task learning (identity + spoof), adversarial training, and spectro-temporal consistency checks for replay/deepfake artifacts. Privacy is addressed using federated learning, keeping raw recordings local and sharing only model updates. Experiments on VoxCeleb2, ASVspoof 2021, and a forensic-style speaker comparison corpus achieve statistically significant performance gains, 98.43% mean identification accuracy with strong class-balanced performance (macro F1 = 98.10%, precision = 98.22%, recall = 98.01%) and statistically significant gains over strong baselines across repeated folds (F1: p=8.0×〖10〗^(-4); precision: p=1.1×〖10〗^(-3); recall: p=9.0×〖10〗^(-4)). The model remains lightweight (≈4.3M parameters, ≈1.2 GFLOPs per 3 s), enabling near real-time inference with modest overhead from consistency checks (<6%). Overall, ForenVoice-Secure provides a compact and reproducible forensic audio data mining framework for scalable, spoof-resilient, privacy-aware law-enforcement identification.

Author 1: Mubarak Albathan

Keywords: Forensic audio data mining; forensic voice analytics; voice biometrics; criminal identification; speaker recognition; anti-spoofing; deepfake and replay detection; convolutional neural networks; long short-term memory; federated learning; privacy-preserving biometrics; law enforcement intelligence systems

Paper 16: GAN-Based Generation of Pre Disaster SAR for Earthquake Interferometry

Abstract: This study proposes an earthquake disaster detection method based on interferometric synthetic aperture radar (InSAR) using synthetic pre‑disaster SAR data generated from optical satellite images. Conventional InSAR analysis requires pre‑ and post‑disaster SAR image pairs acquired under strict orbital and observation constraints, which makes it difficult to obtain suitable pre‑disaster data. In the proposed approach, a digital elevation model (DEM) and land‑cover information are combined with optical imagery, and generative adversarial networks (GANs), specifically pix2pixHD and CycleGAN, are used to generate pseudo‑SAR data that include both amplitude and phase components. Experimental results using Sentinel‑1 SAR and Sentinel‑2 multispectral instrument (MSI) data demonstrate that pix2pixHD achieves higher conversion accuracy than CycleGAN, with a peak signal‑to‑noise ratio (PSNR) of 21.25 dB and a histogram intersection of 65.25%, and that the generated pre‑disaster SAR images can be interfered with post‑disaster SAR observations to detect earthquake‑induced surface changes in the 2024 Noto Peninsula event. These findings indicate that the proposed method can extend the applicability of InSAR to areas and events where suitable pre‑disaster SAR acquisitions are unavailable, contributing to rapid earthquake disaster assessment.

Author 1: Kohei Arai

Author 2: Kengo Ohiwane

Author 3: Hiroshi Okumura

Keywords: GAN; SAR; earthquake; disaster; DEM; pix2pixHD; CycleGAN; interferometric SAR

Paper 17: A Competitive Co-Evolutionary Approach for the Nurse Scheduling Problem

Abstract: The Nurse Scheduling Problem (NSP) is a constrained combinatorial optimisation problem that plays a critical role in healthcare scheduling and constraint optimisation. Traditional evolutionary approaches often rely on static fitness evaluation, which struggles to balance feasibility and solution quality under complex real-world constraints. This study proposes a competitive co-evolutionary algorithm for the NSP that introduces adaptive adversarial evaluation, where candidate schedules are assessed under dynamic competitive pressure to expose structural weaknesses and guide evolution more effectively. The proposed competitive NSP is evaluated on a 20-nurse, one-week scheduling instance and compared against a classical Genetic Algorithm (GA) under identical conditions for 30 independent runs. Experimental results show that the competitive NSP achieves a mean best penalty of 447.28, compared to 651.30 for the classical GA, corresponding to an average improvement of approximately 31%. The competitive approach further exhibits smoother convergence behaviour across generations, indicating stronger optimisation dynamics and improved robustness. These findings demonstrate that competitive co-evolution provides an effective and practical alternative to static fitness-based evolutionary methods for nurse scheduling, with broader applicability to healthcare scheduling and constraint optimisation problems.

Author 1: Maizatul Farhana Mohamad Nazri

Author 2: Zeratul Izzah Mohd Yusoh

Author 3: Halizah Basiron

Author 4: Azlina Daud

Keywords: Nurse Scheduling Problem; competitive co-evolution; evolutionary algorithms; healthcare scheduling; constraint optimisation; adversarial evaluation

Paper 18: Metaphorical Meaning Integration in Poetry Based on Online Discourse Data: Analysis from a Cognitive Linguistics Perspective

Abstract: As an emerging literary form, online poetry has garnered significant attention due to its rapid dissemination, diverse styles, and complex metaphorical expressions. However, the process of metaphorical meaning integration in poetry is difficult to quantify, necessitating support from Artificial Intelligence technologies. This study integrates cognitive linguistics theory with AI algorithms to propose a three-dimensional fusion analysis framework—“cognitive theory + specific AI algorithms + online discourse data”—for dissecting metaphorical meaning integration in online poetry. By constructing a comprehensive methodology encompassing metaphor identification, semantic mapping, and integration analysis, this study offers a novel quantitative pathway for metaphor research in poetry. Experimental validation demonstrates that the integrated approach—leveraging Support Vector Machines (SVM), Convolutional Neural Networks (CNN), BERT pre-trained models, and the DeepSeek-R1 large model—achieves outstanding performance in metaphor recognition accuracy, semantic association quantification, and fusion effectiveness evaluation, fully embodying both theoretical and practical value.

Author 1: Ying LIU

Author 2: Jiting XUE

Keywords: Online discourse data; poetic metaphor; cognitive linguistics; Artificial Intelligence; semantic fusion

Paper 19: Dynamic Decision Model for Tunnel Cross-Passage Layout Based on Multi-Source Sensor Data Fusion

Abstract: The layout of tunnel cross-passages is a critical aspect of tunnel construction and operational safety. Traditional methods, primarily based on static design, struggle to adapt to complex and variable geological and construction environments. This study proposes a dynamic decision model for cross-passage layout based on multi-source sensor data fusion to enhance the scientific rigor and adaptability of cross-passage design. A three-dimensional data fusion mechanism integrating “temporal-spatial-statistical” dimensions was developed. Bayesian network quantifies uncertainty, Kalman filter processes time series data, and PCA extracts spatial features. Reinforcement learning and non-dominated sorting genetic algorithm II (NSGA-II) are used to achieve multi-objective optimization of safety coverage and construction efficiency. The proposed model significantly outperforms the traditional methods in many indicators, and is verified by 100 Monte Carlo simulations and actual tunnel experiments. The dynamic scheme increased the safety coverage rate from 72.4% to 91.7%, shortened the average evacuation distance by 38.7% (from 248 meters to 152 meters), saved resources by 14.2% (about 9.8 million yuan), and shortened the construction period by 3-6 days. The comprehensive utility value is 0.91, which is 19% higher than the traditional static method, and the robustness is enhanced. The model realizes the safe, economical, and efficient real-time optimization of the layout of the transverse channel. It provides a technical path and data support that can be promoted for intelligent tunnel construction under complex geological conditions.

Author 1: Xuejun Di

Author 2: Musha Ruzi

Author 3: Angang Liu

Keywords: Multi-source sensor data fusion; dynamic cross-passage deployment decision-making for tunnels; Kalman filter; reinforcement learning; non-dominated sorting genetic algorithm II

Paper 20: An AI-Driven VR Learning Framework Using RL-Optimized Transformer Models for Personalized English Proficiency Assessment

Abstract: Effective English language learning demands adaptive, interactive, and flexible instructional support, which traditional e-learning systems and existing AI tutors struggle to provide due to limited immersion, static feedback mechanisms, isolated task structures, and the absence of robust reward-driven learning strategies. Although prior studies on VR-based learning environments and Natural Language Processing (NLP) have reported enhanced learner motivation and engagement, most existing solutions suffer from fixed task sequencing, limited real-time linguistic intelligence, and inadequate grammar and pronunciation correction capabilities. To address these challenges, this study proposes a Virtual Reality–based architecture named the Self-Evolving Neural Intelligence Tutor (SENIT), driven by Curriculum Reinforcement Learning and Hierarchical Adaptive Weighting. SENIT integrates a fine-tuned T5 transformer for grammar refinement and prosody-aware feedback, while a reinforcement learning agent dynamically adjusts task difficulty and lesson progression based on learner performance. Developed using Python and TensorFlow and deployed within a Unity3D VR environment, SENIT enables realistic conversational simulations and multimodal learner assessment. Experimental evaluation on a dedicated VR English Learning Dataset demonstrates grammar and pronunciation accuracy improvements of 90% and 81%, respectively, outperforming existing models by approximately 12 percentage points. Additionally, learners achieved notable fluency gains and high engagement scores, highlighting SENIT’s effectiveness in delivering personalized, immersive language learning experiences.

Author 1: A. Sri Lakshmi

Author 2: E. S. Sharmila Sigamany

Author 3: Revati Ramrao Rautrao

Author 4: K. Ezhilmathi

Author 5: Dr. Bhuvaneswari Pagidipati

Author 6: Elangovan Muniyandy

Author 7: Dr. Adlin Sheeba

Keywords: AI-driven learning; Virtual Reality; English language education; reinforcement learning; Natural Language Processing

Paper 21: Advanced Multimodal AI for Resilient Healthcare: Enhancing Early Risk Assessment in Critical Care

Abstract: This study develops an advanced multimodal AI framework to strengthen early risk assessment in critical care and support resilient healthcare delivery. Utilizing the MIMIC-III database, this research extracted structured variables and clinical notes from 26,829 adult patients. A text mining approach based on the BERTopic model was employed to generate topic embeddings from unstructured notes, which were subsequently integrated with 16 quantitative variables. Six machine learning models: Adaboost, Gradient Boosting, Support Vector Classification (SVC), Bagging, Logistic Regression, and MLP Classifier were trained to predict short-term and long-term mortality outcomes. Model performance was evaluated through AUROC, accuracy, recall, precision, and F1-score metrics. The results demonstrate that integrating topic embeddings with structured data significantly improved short-term risk prediction. The SVC model, in particular, achieved an AUROC of 0.9137 for predicting 2-day mortality. Critical predictors identified included the Glasgow Coma Scale, White Blood Cell Count, and text-derived topics related to cardiovascular and neurological conditions. The study is based on a single-center dataset, limiting generalizability. Additionally, only a subset of textual data sources was analyzed, and improvements in long-term risk prediction were relatively modest. These findings demonstrate how multimodal AI can significantly improve early risk assessment and enhance resilience in critical care decision-making. This research pioneers the integration of BERTopic-based text mining with machine learning models for clinical risk prediction, highlighting the value of multimodal data fusion in improving predictive accuracy and enriching medical informatics.

Author 1: Shih-Wei Wu

Author 2: Chengcheng Li

Author 3: Te-Nien Chien

Author 4: Yao-Yu Zhang

Keywords: Resilient healthcare; multimodal AI; early risk assessment; critical care; clinical text mining

Paper 22: Improved Sparrow Search Algorithm-Based Recurrent Neural Network for Short-Term Generation Load Forecasting of Hydropower Stations

Abstract: To address the challenges of low accuracy and high randomness in short-term hydroelectric load forecasting within Multi-energy Coupled Virtual Power Plants (MC-VPPs), this study proposes a hybrid model integrating Variational Mode Decomposition (VMD), Long Short-Term Memory (LSTM) networks, and an Improved Sparrow Search Algorithm (ISSA). Traditional methods, such as exponential smoothing and multiple linear regression, often fail to capture nonlinear dynamics and external disturbances. The proposed framework first decomposes raw load data into four intrinsic mode functions (IMFs) via VMD to extract multi-scale features, including long-term trends, seasonal cycles, and short-term fluctuations. LSTM networks are then applied to model the temporal dependencies of each IMF. To enhance optimization, ISSA introduces a bidirectional sine-cosine search strategy, balancing global exploration and local exploitation to avoid premature convergence. Validated on 1,247 daily load records from a hydropower station in southwestern China, the ISSA-VMD-LSTM model achieves a 30.2% improvement in R², with reductions of 47.2% in RMSE, 47.8% in MAE, and 63.3% in MAPE, outperforming benchmarks like PSO-LSTM and SSA-VMD-LSTM. This demonstrates its robustness in handling nonlinearity and stochasticity. The model enhances MC-VPPs’ operational efficiency by enabling intelligent scheduling and renewable energy integration, with future applications extending to real-time forecasting and other renewable energy systems.

Author 1: Liyuan Sun

Author 2: Yilun Dong

Author 3: Junwei Yang

Keywords: Power plant; load forecasting; Mode Decomposition; Long Short-Term Memory; Sparrow Search Algorithm

Paper 23: Integrating Augmented Reality Learning Objects in Intelligent Tutoring Systems: A Conceptual Model for Engaging Learning Experiences

Abstract: Despite the progress achieved by E-learning platforms, several limitations remain in sustaining learner engagement over time. With the rapid evolution of information and communication technologies, augmented reality has emerged as a powerful medium for designing pedagogical objects that are interactive, immersive, and adaptable to diverse learning contexts. The integration of augmented reality- based learning objects into intelligent tutoring systems enhances the educational process by providing learners with contextualized, multisensory experiences that align with their preferences and profiles. In this perspective, our objective is to propose a model for augmented reality-based learning objects within the context of an Intelligent Tutoring System. The proposed framework addresses a critical research gap: the absence of systematic architectural models that enable real-time, bidirectional adaptation between AR content representation and ITS decision-making mechanisms. Our model aims to strengthen learner motivation and reduce the risk of disengagement by dynamically adapting content to individual needs. It provides a structured foundation for the design and development of augmented reality-based learning objects within an intelligent tutoring system, ensuring that immersive resources are not only technologically innovative but also pedagogically aligned and personalized through the system’s diagnostic and feedback capabilities.

Author 1: Hind Tahir

Author 2: Najoua Hrich

Author 3: Salma El Boujnani

Author 4: Mohamed Khaldi

Keywords: Augmented reality; intelligent tutoring systems; E-learning; adaptation; engagement; personalized learning; learning objects

Paper 24: Comparative Evaluation of CNN Architectures for Corn Leaf Diseases Classification

Abstract: Corn has particular importance in the global food industry. Many diseases attack the corn crops, which affects the crop yield. Early classification and detection of these diseases are pivotal to preventing damage and achieving high crop productivity. Although deep learning, especially convolutional neural networks, has accomplished remarkable results in image recognition, selecting the optimal architecture and using limited datasets remains a challenge. To address this gap, a transfer learning approach based on ImageNet weights was applied to classify three common corn diseases (i.e., gray spot, common rust, and blight), as well as the healthy plants. Six CNN architectures—DenseNet201, EfficientNetB0, VGG16, ResNet50, InceptionV3, and InceptionResNetV2—inclusive performance was evaluated for classification on a corn dataset. Based on evaluation metrics, EfficientNetB0 achieves the highest training accuracy of 97.67% with a fast computational time of 71 seconds. It performs more efficiently than the other architectures. These findings support the use of deep learning models, particularly EfficientNet, in the evolution of artificial intelligence image classification system applications.

Author 1: M. Abdallah

Author 2: M. F. Abu-Elyazeed

Keywords: Artificial intelligence; Convolutional Neural Network; deep learning; image processing; corn diseases classification

Paper 25: Attention-Enhanced Hierarchical Transformer for Multimodal Integration of Mammograms and Clinical Data

Abstract: Breast cancer has been listed as one of the leading causes of death amongst women all over the world, and the current diagnostic techniques, which are founded on the manual examination of mammograms or individual clinical presentations, are often subjective, neither being consistent nor generalizable. The existing computer-aided diagnosis (CAD) systems are also characterized by significant weaknesses related to poor multimodal integration, no interpretability, and vulnerability to class imbalance. In order to address the inadequacy, the present study introduces an advanced multimodal deep learning framework named Hybrid Graph-Generative Transformer (HGGT), designed to integrate high-resolution mammographic images with the clinical, demographic, proteomic, and histological data pertinent to the patient. The HGGT network is a hierarchical Swin Transformer and CNN-based feature extraction, a Graph Attention Network (GAT) (to identify clinical variable interaction), and a contrastive cross-modal generative fusion system (to match the different modalities). The diagnostic head employs a Bayesian uncertainty-aware classifier to ensure more reliability in the prediction of malignancy. It is trained on 5-fold cross-validation, AdamW, and a cosine annealing scheduler, which is set on Python 3.10. It is demonstrated by the performance of the CBIS-DDSM mammography dataset and a corresponding clinical dataset consisting of over 400 patients that HGGT is much superior with 98.2% accuracy, 98.7% precision, 98.5% recall, 99.2% F1-score, and 99.1% AUC-ROC, having a significant advantage over the established models of ResNet50, EfficientNet-B0 and GAN-enhanced CNN classifier. Overall, the HGGT framework is delivering a scalable, interpretable, and highly accurate diagnosis solution that was a huge improvement over the existing unimodal and poorly integrated CAD system in the detection of breast cancer.

Author 1: N. Kannaiya Raja

Author 2: V S Krushnasamy

Author 3: Nurilla Mahamatov

Author 4: Prasad Devarasetty

Author 5: S.T. Gopukumar

Author 6: Sanjiv Rao Godla

Author 7: Vuda Sreenivasa Rao

Keywords: Breast cancer diagnosis; multimodal deep learning; Graph Attention Network; Bayesian uncertainty estimation; explainable AI

Paper 26: Health Information Exchange in Malaysia: Leveraging Interoperability on International Standards for Health Data Exchange

Abstract: This research proposal investigates the implementation of Health Information Exchange (HIE) in Malaysia, focusing on understanding the technological, policy, and financial challenges and opportunities associated with its adoption and effectiveness. Through a comprehensive survey approach targeting healthcare practitioners, policymakers, IT professionals, and patients, the study aims to elucidate the current state of HIE, assess interoperability with international standards, and identify pathways for enhancement. Key objectives include analyzing technological barriers, evaluating policy and regulatory impacts, and exploring sustainable financial models for HIE. Findings indicate a positive trajectory towards HIE implementation, underscored by a broad recognition of its potential to transform healthcare delivery. However, challenges such as system integration, policy clarity, infrastructural readiness, and privacy concerns remain. Recommendations for future improvement emphasize strengthening infrastructure, clarifying policies, enhancing security measures, providing continuous training, fostering innovation, and increasing patient engagement. Furthermore, this study highlights the alignment of HIE implementation with the Sustainable Development Goals (SDGs), particularly SDG 3, to ensure universal health coverage and enhance the healthcare workforce's capacity for process innovation. Incorporating international best practices and a validated framework will further strengthen Malaysia’s healthcare system in the digital age.

Author 1: Mohd Noor A. M. N. L

Author 2: M. Batumalay

Author 3: Balasubramaniam Muniandy

Author 4: Lakshmi D

Author 5: Vinoth Kumar P

Keywords: Health data integration; interoperability; international health data standards; implementation; policy reforms; process innovation

Paper 27: Privacy-Preserving Adaptive Biometric Framework with Reinforcement Learning and Blockchain-Enabled Multi-Factor Authentication

Abstract: Ensuring secure and privacy-preserving authentication in web applications remains a critical challenge due to the limitations of conventional single-factor approaches, which are vulnerable to attacks and fail to account for dynamic user behaviors. Existing multi-factor authentication (MFA) methods often rely on static rules, exposing users to unnecessary friction or weak security under evolving threat conditions. To address these gaps, this study proposes PPAB-RL, a Privacy-Preserving Adaptive Biometric framework leveraging Reinforcement Learning for intelligent MFA selection. The proposed method integrates homomorphic encryption for secure fingerprint feature storage, contextual risk scoring based on device, behavioral, and geolocation deviations, and RL-driven adaptive MFA to dynamically select authentication pathways from password-only to multi-step biometric verification. Implementation is carried out using Python, with biometric processing performed on the SOCOFing dataset containing 6,000 fingerprint images, and blockchain-enabled logging for immutable and tamper-proof audit trails. Experimental results demonstrate that PPAB-RL achieves 96.8% authentication accuracy, surpassing traditional password-only (84.2%) and fingerprint-only (93.5%) methods, while maintaining low encrypted matching overhead and minimal user friction. Ablation studies confirm the essential contribution of each module, biometric preprocessing, encryption, risk analysis, and RL-based adaptation to overall system robustness. The RL policy converges rapidly, allowing real-time adaptation to changing user behaviors and threat contexts. Overall, the proposed PPAB-RL framework establishes a highly secure, intelligent, and scalable authentication paradigm, combining encrypted biometrics, dynamic risk assessment, and blockchain validation, offering an innovative approach that can inspire further research in next-generation privacy-sensitive authentication systems.

Author 1: P. Selvaperumal

Author 2: Sakshi Malik

Author 3: Asfar H Siddiqui

Author 4: Dekhkonov Burkhon

Author 5: Elangovan Muniyandy

Author 6: Garigipati Rama Krishna

Author 7: P N V Syamala Rao M

Keywords: Privacy-preserving authentication; multi-factor authentication; reinforcement learning; biometric verification; blockchain-enabled logging

Paper 28: Intelligent Fruit-Picking Robot Using Convolutional Vision and Kinematic Control for Automated Harvesting

Abstract: This study presents the design, development, and evaluation of an intelligent fruit-picking robot that integrates convolutional vision, adaptive gripping mechanisms, and kinematic control to enable automated harvesting in diverse orchard environments. The proposed system combines a dual-manipulator platform with an extendable scissor-lift mechanism to achieve wide workspace coverage, allowing efficient access to fruits located at varying canopy heights. A deep learning-based recognition module, trained on a Mixed Fruit Dataset, is employed to detect and classify fruits under challenging conditions characterized by occlusions, variable illumination, and dense foliage. Visualization of feature activations confirms that the model effectively focuses on discriminative fruit regions, supporting precise alignment of the end-effector during grasping. The adaptive gripper, designed with compliant materials and multi-configuration geometry, ensures gentle handling across fruits of different shapes and sizes, minimizing mechanical damage. Experimental evaluations demonstrate that the system performs reliably across multiple fruit species, achieving accurate identification, robust segmentation, and stable manipulation in real-field scenarios. The integrated results highlight the robot’s potential to reduce labor dependency, improve harvesting efficiency, and support scalable automation in mixed-crop orchards. Future work will address enhancements in real-time processing, autonomous navigation, and cross-species generalization to advance fully autonomous orchard operations.

Author 1: Nurbibi Sairamkyzy Imanbayeva

Author 2: Bekzat Ondasynuly Amanov

Author 3: Aigerim Bakatkaliyevna Altayeva

Author 4: Dana Kairatovna Ashimova

Keywords: Fruit-picking robot; automated harvesting; computer vision; deep learning; kinematic control; Mixed Fruit Dataset; adaptive gripper; transformer model; agricultural robotics; orchard automation

Paper 29: Ghost-Vanilla Feature Maps: A Novel Hybrid Architecture for Efficient Fine-Grained Songket Motif Classification

Abstract: South Sumatra songket motifs present a challenging fine-grained classification task due to high inter-class similarity and substantial intra-class variability. This study proposes the Ghost-Vanilla Feature Map, a novel hybrid architecture that integrates low-cost ghost-generated features with the lightweight structural stability of VanillaNet to enhance discriminative feature learning while reducing computational burden. The proposed architecture is designed to address the inefficiency of conventional convolution-heavy networks in capturing subtle motif variations. Experimental evaluation on a dataset comprising 20 songket motif classes demonstrates that a ghost ratio 2 achieves the best trade-off, attaining an accuracy of 0.98 with more than 75% parameter reduction. Increasing the ghost ratio to 3 preserves high classification performance with an accuracy of 0.97, while ratios 4 and 5 further reduce model size at the expense of marginal accuracy degradation. Comparative results indicate that the Ghost-Vanilla Feature Map consistently outperforms lightweight CNN baselines, including MobileNetV3-Small, MobileNetV4-Conv-Small, EfficientNetV2-Small, and ShuffleNetV2. The proposed architecture substantially surpasses the Vanilla-only baseline, which achieves an accuracy of only 0.860 despite requiring 30.19 million parameters, highlighting the limitations of conventional convolution-dominant designs in fine-grained textile classification. The hybrid configuration with a ghost ratio 2 delivers superior accuracy while nearly halving the parameter count and significantly reducing computational overhead. Overall, the Ghost-Vanilla Feature Map provides an efficient and highly discriminative solution for fine-grained songket motif classification, achieving strong performance while substantially reducing model complexity through a balanced hybrid representation.

Author 1: Yohannes

Author 2: Muhammad Ezar Al Rivan

Author 3: Siska Devella

Author 4: Tinaliah

Keywords: Ghost Module; fine-grained classification; lightweight deep learning; songket motif classification; VanillaNet

Paper 30: Forecast of Guangzhou Port Logistics Demand Based on Back Propagation Neural Network

Abstract: In recent years, the port economy of our country has developed rapidly. Guangzhou Port is an important node of the maritime transportation of the Belt and Road, connecting the hinterland economy of our country with the countries along the Belt and Road, which is of great significance in promoting the economic development of the hinterland of our country. It is of great significance to predict the freight development demand of Guangzhou port scientifically and reasonably, which is beneficial to optimize the infrastructure construction and logistics system planning of Guangzhou port. This study selects the port cargo throughput, foreign trade cargo throughput, and container cargo throughput as three index values to measure the freight development of Guangzhou port. Firstly, the GM(1,1) model and the BP neural network model are constructed to predict the freight demand of Guangzhou port. Then, the GM(1,1) model and the BP neural network model are combined to predict again. By comparing the three models, the results show that the accuracy of the combined model is better than that of the single model. The combined model of BP neural network and GM(1,1) can be effectively applied in the prediction of Guangzhou port logistics demand. Finally, the combined model of BP neural network and GM(1,1) is used to forecast the freight development demand of Guangzhou Port in 2022-2024, which provides a reference for the development planning of Guangzhou Port. The results further indicate that the BP–GM(1,1) combination model significantly outperforms single forecasting models in terms of prediction accuracy, highlighting its effectiveness and robustness in port logistics demand forecasting.

Author 1: Xiu Chen

Author 2: Lianhua Liu

Author 3: Lifen Zheng

Keywords: BP neural network; GM(1,1); combination model; port logistics demand

Paper 31: Explainable AI Techniques for Interpretable Breast Cancer Classification

Abstract: Breast cancer is still a major health risk for women all over the world, and thus finding it early is very important for the patient's survival. Digital Breast Tomosynthesis (DBT) offers enhanced imaging capabilities relative to conventional mammography; yet, its quasi-3D characteristics provide distinct interpretability issues, often rendering deep learning models as black boxes. This work tackles the issue of transparency by testing three Explainable Artificial Intelligence (XAI) methods: Gradient-weighted Class Activation Mapping (Grad-CAM), Score-CAM, and Local Interpretable Model-Agnostic Explanations (LIME). The ResNet-50 architecture was utilized to analyse a dataset of 396 DICOM images that had been pre-processed in a unique way, including colour-mapping and balancing. The study used Insertion and Deletion Area Under the Curve (AUC) to carefully quantify how reliable the visual explanations were, in addition to usual criteria like accuracy, which achieved 94%. It was shown that LIME and Score-CAM generated attention maps that were dispersed or inconsistent, whereas Grad-CAM always showed lesion-specific areas with great accuracy. Grad-CAM was the best method for analysing DBT findings, since it had the highest Insertion AUC of 0.9078. These results provide radiologists with a way to trust and check automated diagnoses, which closes the gap between AI that works well and AI that is reliable in the clinic.

Author 1: Tony K. Hariadi

Author 2: Qodri Aziz

Author 3: Slamet Riyadi

Author 4: Kamarul Hawari Ghazali

Author 5: Khairunnisa Binti Hasikin

Author 6: Tri Andi

Keywords: Breast cancer; DBT; Grad-CAM; ResNet-50; XAI

Paper 32: Multi-Omics Integration Methods for AI-Based Breast Cancer Molecular Subtypes Classification

Abstract: Breast cancer is one of the most life-threatening and heterogeneous diseases. It contains various molecular subtypes, each subtypes have different characteristics, treatment outcomes, and prognosis. The proper integration of multi-omics data, including genomics, epigenomics, transcriptomics, and proteomics, is very important for enhancing the breast cancer molecular subtypes classification accuracy. Despite the increase in high-dimensional multi-omics data, selecting a suitable integration method for multi-omics data in breast cancer molecular subtypes classification still remains a crucial challenge. This study aims to evaluate and compare, and assess the effectiveness of the multi-omics data integration methods, including exploring the advantages, limitations, and highlighting their performance in terms of accuracy, interpretability, scalability, and biological relevance. Our findings indicate that transformer-based integration methods are increasingly adopted in recent studies due to their superior ability to handle high-dimensional heterogeneous data and capture intricate cross-omics relationships while providing interpretable insights. Additionally, we provide a comparative overview of existing models, discuss key trends over the years, and offer actionable guidance for method selection based on dataset characteristics and research objectives. Finally, we suggest future research directions, emphasizing hybrid deep learning frameworks, graph-based models, and attention mechanisms to enhance predictive accuracy and biological interpretability.

Author 1: Sajid Shah

Author 2: Azurah A Samah

Author 3: Siti Zaiton Mohd Hashim

Author 4: Sarahani Harun

Author 5: Zuraini Binti Ali Shah

Author 6: Farkhana Binti Muchtar

Author 7: Syed Hamid Hussain Madni

Keywords: Breast cancer; classification; integration methods; molecular subtypes; multi-omics

Paper 33: Environmental Assessment of Chemicals: Artificial Intelligence for Predicting Persistence, Bioaccumulation, and Toxicity Properties

Abstract: Early assessment of the persistence, bioaccumulation, and toxicity (PBT) of chemicals is a major challenge for environmental protection and international regulatory frameworks. The objective of this study is to compare the effectiveness of three graph-based deep learning architectures—a graph neural network (GNN), a message passing network (MPNN), and a graph attention network (GAT)—for the binary classification of molecules as PBT or non-PBT.We compiled a regulatory dataset comprising 5,130 molecules annotated from public sources, such as ECHA and international POP lists. Molecular graphs were generated from SMILES using RDKit. The three models were implemented in PyTorch Geometric with homogeneous hyperparameters. The experiments were conducted with a scaffold split ratio of 80/10/10 and 10-fold cross-validation. Performance was evaluated using accuracy, AUC-ROC, and F1-score. Interpretability was examined using GAT model attention maps and atomic contribution analysis. The MPNN model achieves the best overall performance (Accuracy = 0.92; ROC-AUC = 0.94; F1 = 0.91), followed by GAT (Accuracy = 0.89; ROC-AUC = 0.93). The basic GNN performs less well (Accuracy = 0.82; ROC-AUC = 0.89). The GAT model provides more detailed atomic explanations thanks to attention weights, while the MPNN stands out for its predictive accuracy. The dataset includes annotations from heterogeneous experimental sources, which may introduce noise into the labels. The models rely solely on 2D graphs, without 3D conformational information. MPNN models can accelerate PBT pre-screening and help prioritize substances for experimental testing. GATs provide useful interpretations for understanding the substructures associated with PBT properties. This study provides the first reproducible and systematic comparison of GNN, MPNN, and GAT models applied to a large regulatory dataset dedicated to PBT, analyzing both performance and interpretability. These results highlight the potential of graph-based QSAR models for regulatory PBT screening and environmental risk assessment.

Author 1: Ayoub Belaidi

Author 2: Rachid El Ayachi

Author 3: Mohamed Biniz

Author 4: Mohamed Oubezza

Author 5: Youssef Youssefi

Keywords: PBT prediction; persistence; bioaccumulation; toxicity; QSAR models; cheminformatics; environmental risk assessment

Paper 34: A Review on Intrusion Detection Models in Internet of Medical Things (IoMT)

Abstract: The Internet of Medical Things (IoMT) environment is highly sensitive due to the nature of medical data and its direct connection to patient health, making it a prime target for sophisticated cyberattacks. This study explores the key security challenges within IoMT, discusses how Machine Learning (ML) can enhance threat detection capabilities, and shows how XAI contributes to improving transparency and understanding of model decisions, thereby increasing trust in these systems. It reviews recent advancements in Intrusion Detection Systems (IDS) specifically designed for IoMT networks, with a focus on integrating Explainable Artificial Intelligence (XAI) and ML models. Furthermore, the study compares various algorithms and models, identifying research gaps and discussing different datasets and feature extraction techniques used for optimizing the features. The reported performance and efficiency improvements are derived from prior studies using different dataset sizes, data-splitting strategies, and feature-selection methods.

Author 1: Aljorey Alqahtani

Author 2: Monir Abdullah

Keywords: Internet of Medical Things (IoMT); IDS; Explainable Artificial Intelligence (XAI)

Paper 35: A Blockchain-Based Privacy-Preserving Scheme for Integrity Verification and Fair Payment in Cloud Data Storage

Abstract: Ensuring the integrity of outsourced data in cloud storage remains a critical challenge, especially when existing auditing schemes rely on centralized third-party auditors (TPAs), which introduce single points of failure, privacy leakage risks, and a lack of economic fairness. Current blockchain-based approaches improve transparency but still fail to simultaneously achieve privacy-preserving verification and fair payment between data owners and cloud service providers (CSPs). To address this gap, this study proposes a blockchain-based integrity verification scheme that supports decentralized, privacy-preserving, and economically fair audits for encrypted cloud data. The proposed scheme integrates homomorphic linear authenticators (HLA) and multi-party computation (MPC) to verify data integrity without revealing plaintext, while smart contracts are used to enforce automatic payment or penalty based on audit results, ensuring fairness and accountability. A prototype implementation confirms the practicality of the system. Experimental results show that the audit latency is reduced by up to 35 per cent and smart contract gas consumption by approximately 30 per cent compared to existing schemes, while maintaining low computation and communication overhead. Security analysis demonstrates that the scheme provides data integrity, privacy protection, fairness, and resistance to replay and collusion attacks. Overall, this work offers a practical and scalable solution for secure cloud storage auditing.

Author 1: Li Zhenxiang

Author 2: Jin Yuanrong

Author 3: Mohammad Nazir Ahmad

Keywords: Cloud storage; blockchain; integrity verification; smart contract; privacy-preserving audit; fair payment

Paper 36: EEG-Based Imagined-Speech Decoding: A Review

Abstract: Non-invasive neural speech interfaces aim to reconstruct intended words from brain activity, offering critical communication options for individuals with severe dysarthria or locked-in syndrome. Among the available recording modalities, electroencephalography (EEG) remains the most accessible and cost-effective choice for long-term brain–computer interface (BCI) applications. Decoding imagined speech from EEG, however, remains difficult because of low signal-to-noise ratio, pronounced inter-subject variability, and the small, heterogeneous corpora that are currently available. This review adopts a narrative methodology to synthesise peer-reviewed studies on EEG-based imagined-speech decoding. Relevant articles were identified through keyword-based searches in major digital libraries and were included if they used non-invasive EEG, explicitly instructed imagined or covert speech, and reported quantitative decoding performance. The selected studies are organised along the processing pipeline, from experimental paradigms and data acquisition to preprocessing, feature extraction, representation learning, and classification. Across this body of work, binary imagined-speech tasks that rely on carefully designed time–frequency features and shallow classifiers often report accuracies above 80 percent, whereas multi-class word or phoneme recognition exhibits a much wider spread of performance and remains highly sensitive to dataset design and evaluation protocol. Recent trends favour convolutional and recurrent neural networks, temporal convolutional networks, and transfer learning strategies, which improve performance on some datasets but do not yet resolve fundamental issues of restricted vocabularies, inconsistent evaluation practices, and limited cross-subject generalisation. The review distils these observations into practical recommendations for dataset construction, model design, and evaluation protocols and outlines research directions aimed at more robust and clinically meaningful EEG-based imagined-speech BCIs.

Author 1: Hatem T M Duhair

Author 2: Masrullizam Bin Mat Ibrahim

Author 3: Mazen Farid

Author 4: Jamil Abedalrahim Jamil Alsayaydeh

Author 5: Safarudin Gazali Herawan

Keywords: Electroencephalography (EEG); Imagined speech; Brain–Computer Interfaces (BCIs); neural speech decoding; deep learning; transfer learning; time–frequency analysis; evaluation protocols

Paper 37: Evaluating Field Flexibility Approaches in Relational Databases: A Performance Study of JSON and Column-Oriented Models in Library Systems

Abstract: This study examines two approaches for achieving field flexibility in library systems using relational databases: column-oriented tables and JSON data types. To evaluate the performance and practicality of flexible schema strategies, a dataset of 41,000 library records was implemented using both column-oriented and JSONB-based schemas in PostgreSQL. Five representative queries based on typical search operations in library applications were executed repeatedly on each model, and average execution times were measured in a controlled environment. Results show that JSONB consistently outperforms the column-oriented approach across all query scenarios, benefiting from reduced structural overhead and more direct access to semi-structured data. However, the flexibility of JSONB introduces risks of inconsistent data structures and reduced schema enforcement compared to the more rigid but uniform column-oriented method. The findings highlight a trade-off between performance and data consistency, suggesting that JSONB is advantageous for dynamic, metadata-rich systems, while column-oriented storage remains preferable when strict structural integrity is required. Future work should explore hybrid models and schema validation layers to combine flexibility with reliable data governance.

Author 1: Rizal Fathoni Aji

Author 2: Nilamsari Putri Utami

Keywords: Field flexibility; RDBMS; column-oriented model; JSON; library systems

Paper 38: Developing a Robotic-Integrated Leagility Adaptation Model Through Green Supply Chain Intelligence and Supply Chain Ambidexterity

Abstract: This study develops a Robotic-Integrated Leagility Adaptation Model by combining Green Supply Chain Intelligence (GSCI) and Supply Chain Ambidexterity (SCA) to enhance sustainable performance in the manufacturing sector. The rapid evolution of robotics, cyber-physical systems, and AI-enabled decision technologies has transformed supply chain dynamics, necessitating an adaptive model that balances efficiency (lean) and responsiveness (agile). Using an integrated quantitative approach, this research examines how robotic automation strengthens leagility capabilities through real-time analytics, predictive intelligence, and environmentally oriented digital operations. The findings demonstrate that GSCI significantly enhances SCA, which in turn improves leagility adaptation and sustainable manufacturing performance. Robotic integration is found to play a catalytic role by enabling autonomous coordination, energy-efficient scheduling, and intelligent material handling as key enablers of green and responsive operations. This study contributes to the literature by proposing a technology-driven leagility model that links robotics, green supply chain intelligence, and ambidexterity within a unified smart manufacturing framework. Implications are provided for policymakers and industry leaders to accelerate sustainable transformation through robotics-enabled digital ecosystems.

Author 1: Miftakul Huda

Author 2: Mohammad Hatta Fahamsyah

Author 3: Agung Nugroho

Author 4: Arie Indra Gunawan

Author 5: Pepen Komarudin

Author 6: Andrean Bagus Saputra

Keywords: Robotic integration; Green Supply Chain Intelligence; Supply Chain Ambidexterity; leagility adaptation; sustainable manufacturing performance

Paper 39: A Privacy-Conscious Federated Reinforcement Learning Framework for Affect-Aware English Listening

Abstract: The rapid growth of digital English listening platforms has intensified the need for intelligent personalization mechanisms that adapt to learner progression while preserving data privacy. Existing adaptive systems primarily rely on static difficulty scaling or centralized learning architectures, often neglecting learner engagement dynamics and raising concerns about sensitive data exposure. To address these limitations, this study proposes PrivAURAL, a privacy-preserving and affect-aware adaptive English listening framework that models listening instruction as a sequential decision-making problem. The objective is to dynamically personalize listening tasks by jointly considering comprehension performance and engagement trends, without transmitting raw learner data. PrivAURAL integrates HuBERT-based semantic–acoustic representations with affective proxy signals derived from learner behavior and employs a Federated Deep Q-Network to adapt task difficulty, playback speed, and assessment frequency. The model is implemented using PyTorch, HuggingFace speech models, and a simulated federated learning environment with secure aggregation. Experiments conducted on the TED-LIUM dataset demonstrate a 32.7% reduction in Word Error Rate over ten sessions, a 21.9% decrease in task completion time, and an improvement in listening accuracy from 86.1% to 87.3% compared with non–affect-aware baselines. Federated training further ensures stable convergence, while maintaining strict privacy constraints. The results confirm that reinforcement-driven, affect-aware personalization can significantly enhance listening efficiency and engagement, positioning PrivAURAL as a scalable, ethical, and privacy-conscious solution for next-generation digital language learning systems.

Author 1: N. Sreedevi

Author 2: V. Saranya

Author 3: Kama Ramudu

Author 4: M. Madhusudhan Rao

Author 5: Sakshi Malik

Author 6: Elangovan Muniyandy

Author 7: Ahmed I. Taloba

Keywords: Adaptive listening learning; federated reinforcement learning; affective proxy modeling; privacy-preserving AI; HuBERT speech representation; English language learning

Paper 40: Advanced Explainable Hybrid Metaheuristic–Deep Learning Framework for Real-Time Financial Fraud Detection with Temporal Convolutional Analysis

Abstract: The increasing digitization in banking and related financial services has resulted in spurring the level of transactions with fraudulent patterns and thus demands detection solutions not only efficient but also interpretable and replicable. The earlier machine learning approaches, like K-Nearest Neighbors, Decision Trees, and Random Forests, are not efficient in dealing with high-dimensional and sequential patterns in transactions; in addition, they are incapable of modeling time patterns and are not interpretable models. Since there exist drawbacks in earlier approaches, this work introduces an Interpretable Moth-Flame Optimized Temporal Convolutional Network (MFO-TCN) for efficient and interpretable real-time financial fraud detection. The approach is initiated with rigorous data preprocessing tasks like normalization and encoding performed on the Bank Account Fraud (BAF) dataset. Based on the Moth-Flame Optimization (MFO) algorithm, the optimal features of the transactions expressing high discriminative powers are extracted. This is followed by the application of the Temporal Convolutional Network (TCN) technique, which is capable of identifying the sequential patterns of fraud-related activities. For improved transparency and validity, the SHAP explainability technique has been adopted, ensuring better explanations for feature importance and decision-making. The proposed MFO-TCN results in an accuracy of 97.2% with higher values of precision and recall, achieving better results in comparison to classical and ensemble approaches. Moreover, it provides real-time processing for transactions in milliseconds. The above results show that an efficient combination of metaheuristics for feature optimization, along with temporal deep networks, can provide an optimal technique for financial fraud detection systems.

Author 1: Madhu Kumar Reddy P

Author 2: M. N. V Kiranbabu

Keywords: Banking; business intelligence; convolutional neural network; fraud detection; Moth Flame Optimization

Paper 41: A Few-Shot Semantic Meta-Learning Framework with CRF for Skill Entity Recognition in Open Innovation Ecosystems

Abstract: The accelerating pace of digital transformation is reshaping labour-market dynamics, driving the emergence of new competencies, and intensifying the need for scalable skill-intelligence systems within open innovation ecosystems. Yet, research on Indonesian Named Entity Recognition (NER) remains limited for skill-extraction tasks, especially in low-resource contexts where annotated data are scarce and novel skill expressions evolve rapidly. To address this gap, this study contributes to applied Natural Language Processing (NLP) by introducing the Few-Shot Semantic Meta-Learning framework with CRF (FSM-CRF) for Indonesian skill entity recognition, which integrates semantic span representations, episodic meta-learning, and BIO-constrained CRF decoding to enhance prototype stability and entity-boundary precision for complex, multi-token skill expressions. Using the NERSkill.id dataset, the proposed model is evaluated under a 3-way, 10-shot episodic setting and achieves a micro-F1 of 73.84%, outperforming traditional supervised approaches (IndoBERT fine-tuning, BiLSTM-CRF) and existing few-shot baselines. Ablation experiments further demonstrate that semantic span modelling and structured CRF inference play pivotal roles in improving robustness, while meta-learning strengthens adaptability across diverse and evolving skill categories. From an open innovation perspective, this framework offers a data-efficient solution for dynamic competency mapping, reducing dependence on costly annotation pipelines and enabling continuous updates to workforce skill taxonomies. Overall, the findings highlight semantic meta-learning as a promising foundation for next-generation skill-intelligence infrastructures that support AI-enabled innovation management, strategic workforce planning, and evidence-informed policy design.

Author 1: Nurchim

Author 2: Muljono

Author 3: Edi Noersasongko

Author 4: Ahmad Zainul Fanani

Author 5: Deshinta Arrova Dewi

Keywords: Few-shot; Named Entity Recognition; skill intelligence; process innovation; open innovation; Natural Language Processing

Paper 42: Attention-Guided Lightweight MobileNetV2 for Real-Time Driver Drowsiness Classification on Edge-IoT Systems

Abstract: Driver drowsiness is a major cause of traffic accidents, so Edge-IoT platforms with limited resources need to be able to accurately and quickly detect when drivers are drowsy. This study examines attention-guided lightweight CNN design predicated on MobileNetV2 for real-time driver drowsiness detection. The authors compare a SE-enhanced MobileNetV2 to the baseline model and a structurally optimized version that uses Depthwise Separable Convolution (DSC), Bottleneck blocks, and Expansion layers. Experiments on 500 images demonstrate that channel attention enhances feature discrimination, whereas structural optimization yields the most resilient trade-off between accuracy and latency. Statistical validation employing 95% confidence intervals and two-proportion Z-tests substantiates the significance of these enhancements. The proposed models support real-time inference despite their small size (about 2.6 million parameters and 315 million FLOPs). These findings suggest structural optimization is more important than attention mechanisms in designing lightweight CNNs for embedded driver monitoring.

Author 1: Yo Ceng Giap

Author 2: Muljono

Author 3: Affandy

Author 4: Ruri Suko Basuki

Author 5: Harun Al Azies

Author 6: R. Rizal Isnanto

Author 7: Deshinta Arrova Dewi

Keywords: Driver drowsiness detection; Edge-IoT deployment; lightweight convolutional neural networks; process innovation; MobileNetV2 optimization; squeeze-and-excitation attention

Paper 43: Quality of Service and Customer Satisfaction: A Case Study of Call Center Services

Abstract: Quality of service and customer satisfaction have become priority aspects for call center services, especially in a context where their use is becoming more and more frequent. In the district of Los Olivos, Lima, Peru, 60% of users who receive telephone service consider the quality of the service to be deficient, which shows the need to delve into this issue. The objective of the study was to determine the relationship between service quality and customer satisfaction in call center services in a district of Lima. To this end, a non-experimental, quantitative, correlational and cross-sectional approach was used. A questionnaire was applied to 384 clients to measure both variables and their relationship was analyzed using Spearman's correlation. The results show a positive, very strong and significant correlation between service quality and customer satisfaction (r=0.907; p<0.001). Likewise, the dimensions of service quality were significantly related to customer satisfaction: reliability (r=0.850), responsiveness (r=0.618), safety (r=0.473) and empathy (r=0.587). It concludes by highlighting the importance of strengthening the quality of service to improve customer satisfaction and generate benefits for the company. Finally, the need to investigate additional factors that may influence this dynamic is raised.

Author 1: Jhoanna Iveth Santiago Rufasto

Author 2: Sebastián Ramos-Cosi

Author 3: Haslyd Claydiana Ramos Jara

Author 4: Ana Huamani-Huaracca

Keywords: Quality of service; customer satisfaction; call center

Paper 44: Deep Learning for Endometrium Segmentation in Transvaginal Ultrasound: A Systematic Review Towards Receptivity Assessment

Abstract: Deep learning (DL) has become a transformative approach in medical image analysis, offering superior accuracy and automation in image segmentation tasks. In reproductive imaging, transvaginal ultrasound (TVUS) serves as a crucial modality for evaluating the endometrial condition, which plays a critical role in assessing ovarian health. Although many studies have applied deep learning to the segmentation of pathological endometrial conditions, research focusing on non-pathological endometrium segmentation remains critically limited. This study presents a comprehensive review of deep learning methods for endometrium segmentation in TVUS, with a focus on non-pathological conditions, including endometrial thickness measurement, morphology analysis, and endometrium receptivity assessment. Following PRISMA guidelines, research articles published between 2015 and 2025 were identified from major scientific databases. The selected studies were analyzed in terms of image processing methods, deep learning architectures, and performance metrics, such as Dice coefficient, Jaccard index, precision, recall, and Hausdorff distance. Although foundational architectures, such as U-Net and its variants, achieve impressive Dice coefficients (up to 0.977), the results often rely on small and single-center datasets, proving limited generalizability across imaging settings. Recent advancements demonstrate the efficacy of hybrid architectures, such as the Deep Learned Snake algorithm and Transformer-based models like SAIM, in optimizing segmentation precision within noisy transvaginal ultrasound images. This review highlights the lack of attention to non-pathological endometrium segmentation and guides future research directions in self-supervised learning, transformer-based architectures, and interpretable deep learning to achieve robust and clinically applicable models for enhancing endometrium receptivity assessment and supporting ovarian health in assisted reproduction technology.

Author 1: Asma Amirah Nazarudin

Author 2: Siti Salasiah Mokri

Author 3: Noraishikin Zulkarnain

Author 4: Aqilah Baseri Huddin

Author 5: Mohd Faizal Ahmad

Author 6: Ashrani Aizzuddin Abd Rani

Author 7: Seri Mastura Mustaza

Author 8: Huiwen Lim

Keywords: Endometrium segmentation; deep learning; image segmentation; image processing; endometrium receptivity assessment; ovarian health

Paper 45: Advances in Deep Learning for Affective Intelligence: Language Models, Multimodal Trends, and Research Frontiers

Abstract: The accelerated growth of digital content and the increasing presence of emotional expressions, polarized opinions, and toxic behaviors in social media have driven the development of advanced Affective Analysis techniques. This study presents a broad and up-to-date review of recent studies covering Sentiment Analysis, Emotion Recognition, Hate Speech Detection, cyberbullying, and multimodal approaches grounded in deep learning. The review provides a comparative analysis of the architectures employed—including Transformer-based Models, multimodal frameworks, and variants designed for low-resource languages—along with their metrics, performance outcomes, and emerging patterns. The findings reveal a clear consolidation of Transformer-based Models as the dominant standard, significant progress in multimodality for affective interpretation, and growing attention to multilingual models adapted to diverse cultural contexts. Furthermore, persistent challenges are identified, including limitations related to data availability and quality, Explainable AI (XAI), computational efficiency, and robustness in cross-domain generalization. This review synthesizes current trends, limitations, and opportunities in the field, offering a structured perspective that can serve as a reference for researchers and practitioners involved in the development of more accurate, efficient, and culturally responsible affective systems.

Author 1: Diego Andres Andrade-Segarra

Author 2: Juan Carlos Santillán-Lima

Author 3: Miguel Duque-Vaca

Author 4: Fernando Tiverio Molina-Granja

Keywords: Sentiment Analysis; Emotion Recognition; Hate Speech Detection; cyberbullying; deep learning; Transformer-based Models; Multimodal Analysis; multilingual NLP; Explainable AI (XAI); social media

Paper 46: Architecture of an Intelligent Predictive Analytics System for Gas Environment Monitoring Based on Sensor-Series IoT Devices

Abstract: Industrial facilities operating with toxic and explosive gases require continuous monitoring systems capable not only of detecting threshold exceedances but also of anticipating hazardous trends. Conventional IoT-based gas monitoring solutions are primarily limited to real-time data acquisition and alarm triggering, which restricts their ability to prevent incidents proactively. This study presents the architecture of an intelligent predictive analytics system for gas environment monitoring that integrates sensor-series IoT gas analyzers with advanced data analytics. The proposed system is built on domestically developed SENSOR-Mine gas analyzers supporting LoRaWAN and Wi-Fi communication, centralized data storage in MS SQL Server, machine learning–based analytics implemented in Python, and a web-based visualization platform using ASP.NET MVC. Time-series forecasting models and anomaly detection algorithms are jointly employed to analyze gas concentration dynamics and identify potentially dangerous situations at early stages. Experimental validation using carbon monoxide measurements demonstrates the practical applicability of the proposed architecture for industrial safety monitoring. The presented approach provides a scalable foundation for intelligent gas environment monitoring systems aimed at reducing industrial risks and improving worker protection.

Author 1: Anuar Kussainov

Author 2: Gulnaz Zhomartkyzy

Author 3: Rajermani Thinakaran

Keywords: Intelligent system; predictive analytics; IoT; gas analyzer; LoRaWAN; industrial safety; machine learning

Paper 47: A Systematic Literature Review on Organizational Readiness for Artificial Intelligence Adoption Based on the TOE Framework

Abstract: Artificial intelligence (AI) is increasingly being integrated into organizational processes, reshaping how organizations operate, compete, and make decisions. However, despite growing interest, many organizations face challenges in adopting AI effectively due to insufficient readiness. Prior research on organizational AI readiness has produced diverse and sometimes inconsistent conceptualizations, particularly with respect to definitions, readiness factors, and analytical approaches. To consolidate these dispersed insights, this study undertakes a structured review of the literature to synthesize organizational AI readiness factors through the lens of the Technology–Organization–Environment (TOE) framework. The review applies a transparent and replicable screening and selection process, consistent with PRISMA principles, to analyze peer-reviewed journal articles on organizational AI adoption and readiness. Through a multi-stage coding process, 124 readiness-related indicators were identified and subsequently consolidated into 35 factors, which were further synthesized into 12 core readiness themes mapped across the technological, organizational, and environmental dimensions of the TOE framework. The results indicate that organizational AI readiness is not a standalone condition, but a multidimensional and interdependent construct shaped by the alignment of technological capabilities, organizational structures and competencies, and external environmental conditions. By providing a structured synthesis of organizational AI readiness factors, this study clarifies the multidimensional nature of readiness and highlights cross-dimensional interdependencies within the TOE framework. The findings contribute theoretical clarity to the AI readiness literature and offer a consolidated foundation for future empirical studies and practical readiness assessments in organizational settings.

Author 1: Sulistyo Aris Hirtranusi

Author 2: Benny Ranti

Author 3: Widijanto Satyo Nugroho

Author 4: Wisnu Jatmiko

Keywords: Artificial intelligence readiness; organizational readiness; TOE framework; systematic literature review; AI adoption

Paper 48: Volumetric Feature Learning for High-Fidelity Two-Dimensional Dental Cast Image Reconstruction Using Generative Adversarial Networks (GANs)

Abstract: Dentistry is a medical branch that diagnoses and treats oral diseases, helps maintain oral function, and improves oral aes-thetics. Dental casts are three-dimensional models of a patient’s oral tissues that can be used to study oral anatomy, assess oc-clusal relationships, and determine tooth alignment. Traditional-ly, they were made of gypsum, an impression material used to pour into the patient’s mouth molds. Meanwhile, digital ones are three-dimensional models generated virtually using modern digi-tal imaging and intraoral scanners. Unlike physical models, which require a lot of manual work and ample storage space, digital models can be produced rapidly, easily modified, and stored for long-term usage. In this study, we present Denta-RecGAN, a novel approach based on Generative Adversarial Networks (GANs) that maps a two-dimensional dental cast im-age into a volumetric latent space and projects it back into a two-dimensional output. The proposed approach employs a 2D encoder to process dental cast images as input, enabling the extraction of spatial features. The structural depth is modelled, and noise is suppressed using volumetric 3D latent space de-noising models; a 2D decoder then reconstructs a high-quality image. The model is trained under an adversarial learning ap-proach using the IO150K dataset. The proposed architecture achieved Mean Absolute Error (MAE) of 0.0128, 0.0127, 0.0128; Structural Similarity Index Measure (SSIM) of 0.9450, 0.9452, 0.9453; and Peak Signal-to-Noise Ratio (PSNR) of 28.84, 28.85, 28.84?decibels across training, validation, and testing sets. These results demonstrate the effectiveness of volumetric feature learning in enhancing the accuracy of 2D image re-construction and preserving fine structural details.

Author 1: Eman Ahmed Eldaoushy

Author 2: Manal A. Abdel-Fattah

Author 3: Nermeen Ahmed Hassan

Author 4: Mai M. El defrawi

Keywords: Dental image reconstruction; generative adversarial networks; latent space representation; two-dimensional to three-dimensional mapping; volumetric deep learning

Paper 49: Selfdom Enhanced CatBoost Model for Remote Paddy Growth Monitoring and Fertilizer Recommendation in Precision Agriculture

Abstract: Precision agriculture enables data-driven crop monitoring and improved resource utilization. Paddy cultivation requires continuous surveillance and timely fertilizer application because it is sensitive to soil nutrient dynamics, water availability, and climatic conditions. Conventional practices such as manual field inspection and heuristic fertilizer advisory methods are often labor-intensive and subjective, which can reduce decision consistency and contribute to yield variability. To address these limitations, this study proposes a Selfdom Enhanced CatBoost (SECB) framework for remote paddy growth-stage monitoring and fertilizer recommendation. Multispectral remote sensing data collected over multiple seasons are used to compute vegetation indices, including NDVI, GNDVI, RVI, GRVI, and NDRE, to characterize crop vigor and chlorophyll-related variation across growth stages. The proposed SECB improves CatBoost by integrating an Improved Osprey Optimization Algorithm (IOOA) to tune key model parameters, aiming to enhance feature interaction learning and reduce overfitting. In addition, oppositional function-based initialization is applied to improve the exploration capability of IOOA and accelerate convergence. Experimental results show that SECB achieves improved performance over baseline classifiers in terms of accuracy, precision, F1-score, specificity, and AUC. The proposed approach provides reliable growth-stage identification and supports fertilizer recommendations to promote efficient nutrient usage and improved productivity. Overall, the framework offers an automated and scalable decision-support strategy for paddy crop management.

Author 1: Shanmuga Priya S

Author 2: V. Dhilip Kumar

Keywords: Paddy growth monitoring; fertilizer recommendation; Selfdom Enhanced CatBoost Model; Osprey Optimization Algorithm; oppositional function

Paper 50: A Robust Real-Time Multimodal Polynomial Fusion Framework for Sensor-Based Sign Language Recognition Using Flex–IMU Smart Gloves

Abstract: Sign language recognition is a critical component of assistive technologies for individuals with hearing and speech impairments. While vision-based approaches have shown promising performance, their reliability is often affected by illumination variations, occlusions, and background complexity. Wearable sensor–based solutions, particularly smart gloves integrating flex sensors and inertial measurement units (IMUs), provide a more stable alternative by directly capturing hand articulation and motion patterns. However, existing sensor-based methods frequently suffer from temporal instability, noise sensitivity, and limited discrimination among structurally similar gestures, which is especially challenging in Hijaiyah sign language, where many letters differ only by subtle finger configurations. This study proposes a robust real-time Multimodal Polynomial Fusion (MPF) framework for sensor-based sign language recognition using a flex–IMU smart glove, with a specific focus on Hijaiyah gestures as the application domain. The proposed framework applies nonlinear polynomial temporal smoothing within a sliding window to stabilize raw flex–IMU trajectories, followed by multimodal fusion to enhance gesture separability and temporal consistency. A large-scale multimodal dataset comprising 231,000 samples collected from 33 users performing 28 Hijaiyah gesture classes was constructed to enable rigorous subject-independent evaluation. Experimental results obtained from offline testing, session-aware analysis, and real-time streaming scenarios demonstrate that the proposed MPF framework consistently outperforms a baseline approach based on raw normalized signals. The proposed method improves recognition accuracy from 92.42% to 96.32%, while also achieving higher macro-level precision, recall, and F1-score. Furthermore, MPF significantly reduces misclassification rates and improves temporal stability, particularly for fine-grained Hijaiyah gestures with similar structural patterns. These results confirm that the proposed framework provides a robust and reliable solution for real-time wearable sign language recognition and offers practical benefits for Hijaiyah-based assistive communication systems.

Author 1: Dadang Iskandar Mulyana

Author 2: Edi Noersasongko

Author 3: Guruh Fajar Shidik

Author 4: Pujiono

Keywords: Sign language recognition; Hijaiyah sign language; wearable sensors; smart glove; multimodal fusion; polynomial temporal smoothing; real-time recognition

Paper 51: EnGMHE: Enhanced Geometric Mean Histogram Equalization for Low-Light Image Enhancement

Abstract: Low-light image enhancement has been extensively studied, with numerous methods proposed to address this challenge. Among these, Geometric Mean Histogram Equalization (GMHE) emerged as a histogram-based technique specifically designed for enhancing low-light images. Despite its effectiveness, GMHE has notable limitations: it often oversaturates results under specific conditions and amplifies noise, limiting its practical applicability. These shortcomings become particularly pronounced in real-world scenarios where low-light conditions are frequently accompanied by significant noise artifacts. To address these shortcomings, this study introduces EnGMHE, an enhanced version of GMHE. The proposed method consists of three key steps: 1) introducing a novel Gaussian Histogram Equalization (GHE) to improve image contrast and brightness, 2) utilizing GMHE to enhance sharpness and detail clarity, and 3) denoising the enhanced image using a pretrained deep neural network model. Together, these steps offer a more robust solution for low-light image enhancement, balancing contrast improvement, detail preservation, and noise reduction. The experimental results reveal not only the efficiency but also the effectiveness of the proposed model when benchmarked against the state-of-the-art methods.

Author 1: Rawan Zaghloul

Author 2: Hazem Hiary

Keywords: Histogram Equalization; image enhancement; low-light enhancement; denoising; deep learning

Paper 52: Immersive Learning Environment Design of Outdoor Education Space Using Artificial Intelligence Augmented Reality Technology

Abstract: This study presents a technically grounded design and implementation of an AI- and AR-enabled immersive learning environment for outdoor education. Moving beyond conceptual descriptions, the study develops an executable system framework that integrates adaptive navigation and positioning, context-aware virtual tours, task-driven scenario simulation, and real-time feedback mechanisms. Each functional module is explicitly linked to algorithmic implementations, including multi-sensor state estimation, constrained generative scene construction, and reinforcement-based adaptive control, enabling reproducible system behavior in real outdoor settings. A controlled field experiment was conducted using an experimental group and a control group under identical instructional conditions. Quantitative evaluation based on pre–post testing, behavioral logging, and statistical analysis demonstrates that the proposed system achieves statistically significant improvements in learning interest, participation, knowledge mastery, and problem-solving ability. Experimental conditions, data characteristics, and methodological limitations are explicitly reported to support result verification and generalizability. The findings indicate that the proposed immersive learning environment constitutes a validated system-level contribution rather than a purely conceptual framework, offering practical and scientific value for computer science–oriented educational technology research.

Author 1: Chenguang Liu

Keywords: Artificial intelligence; augmented reality; outdoor education space; immersive learning environment; system-level evaluation

Paper 53: A Contextualized Learner-Profiling Transformer Architecture for Adaptive Grammar Error Diagnosis and Instruction

Abstract: Grammatical accuracy is a critical component of English as a Second Language (ESL) learning; however, many learners continue to struggle with recurring errors despite the availability of automated grammar correction tools. Although recent transformer-based models such as BERT, GPT, and T5 have demonstrated strong benchmark performance, existing grammar error correction (GEC) systems remain largely correction-oriented and lack pedagogical flexibility, learner awareness, and explanation-based feedback. To address these limitations, this study proposes an Adaptive Multi-Task T5 (AMT-T5) framework that integrates grammatical error correction, error-type classification, and personalized feedback generation within a unified transformer architecture. The proposed method is designed to actively support learner development by maintaining dynamic learner error profiles and adaptively reweighting attention to provide targeted instructional guidance. AMT-T5 is implemented using Python, PyTorch, and the Hugging Face Transformers library, and trained on the Lang-8 Learner Corpus, which contains authentic ESL learner sentences with expert corrections. Experimental results demonstrate that the proposed model significantly outperforms existing transformer-based baselines, achieving 78.9 BLEU, 90.7 GLEU, 82.6% full-sentence accuracy, and an error reduction rate of 91.2%, representing an approximate 18–22% improvement in grammatical accuracy over prior models. The framework further incorporates Direct Preference Optimization to align corrections with pedagogical expectations and Knowledge Distillation to enable efficient real-time deployment. Overall, the proposed AMT-T5 framework transforms grammar correction from a passive editing task into an adaptive, learner-centered educational process, offering a scalable and effective solution for intelligent ESL grammar learning systems.

Author 1: Bukka Shobharani

Author 2: M Vijaya Lakshmi

Author 3: Kama Ramudu

Author 4: Jasgurpreet Singh Chohan

Author 5: S. Farhad

Author 6: Elangovan Muniyandy

Author 7: Gulnaz Fatma

Author 8: Ahmed I. Taloba

Keywords: Grammar error correction; adaptive feedback; Multi-task Learning; transformer models; ESL learning; personalized language instruction

Paper 54: Text-Driven Early Warning of Supply Chain Risks: A Hybrid Machine- and Deep-Learning Framework for the New Energy Vehicle (NEV) Industry

Abstract: The rapid expansion of New Energy Vehicles (NEVs) has increased the global NEV supply chains' exposure to diverse and interconnected risks. Distributed production networks frequently face disruptions driven by raw material volatility, evolving environmental regulations, customs clearance uncertainty, and geopolitical instability, underscoring the need for effective early-warning systems. To address limitations in existing studies that lack a consistent and interpretable structure for NEV-specific hazards, this study proposes a hybrid NLP-based pipeline for risk text classification and early-warning sender extraction. A curated dataset of 120 NEV-related risk reports published between 2023 and 2025 was collected from Chinese information sources, pre-processed, and annotated according to a six-category risk taxonomy. Classical machine-learning models, including logistic regression, support vector machines, random forest, and XGBoost, were trained using TF-IDF features, while a multilayer perceptron and a BERT model were employed to capture nonlinear patterns and contextual semantics. Classical models were evaluated using five-fold cross-validation, and deep models were assessed on a held-out test set. XGBoost achieved the best classical performance, with accuracy and F1 scores of 0.826 and 0.766, respectively. BERT outperformed all baselines, reaching an accuracy of 0.864 and an F1 score of 0.808. The proposed framework demonstrates a modular and scalable approach.

Author 1: Ma Chaoke

Author 2: S. Sarifah Radiah Shariff

Author 3: Noryanti Nasir

Author 4: Gao Ying

Keywords: New Energy Vehicle (NEV); supply chain risk; natural language processing (NLP); text classification; early warning system; BERT

Paper 55: Explainable CNN-Based Multiclass Household Waste Classification Using Grad-CAM for Smart Waste Management

Abstract: Automated waste classification using computer vision has become essential for improving environmental sustainability and reducing manual sorting effort. This study presents an enhanced waste image classification model based on EfficientNet-B0, trained using a two-stage transfer learning strategy that combines feature extraction and fine-tuning. The proposed approach aims to enhance classification accuracy while maintaining computational efficiency. Experimental evaluations conducted on a heterogeneous multi-class waste dataset demonstrate the superiority of the proposed method. The confusion matrix results indicate a high proportion of correct predictions across most categories, with only minor misclassifications among visually similar classes, such as metal and paper. The model's robustness is further validated through 5-Fold Cross-Validation, which yields an average accuracy of 94.3% with a standard deviation of ±0.007, confirming consistent performance across data partitions. Compared with state-of-the-art CNN architectures, including ResNet50 and DenseNet121, the proposed model achieves the highest accuracy while using the fewest parameters (4.38M), making it suitable for deployment in resource-constrained environments. Additionally, qualitative analysis using Grad-CAM confirms that the model’s decisions are explainable and based on relevant object features. These findings demonstrate that the proposed EfficientNet-B0 model constitutes a reliable, efficient, and interpretable solution for automated waste classification. The model is further evaluated using cross-validation and explainable AI (Grad-CAM) to assess both performance stability and interpretability.

Author 1: Fuzy Yustika Manik

Author 2: Pauzi Ibrahim Nainggolan

Author 3: T. H. F Harumy

Author 4: Dewi Sartika Br Ginting

Author 5: Aini Maharani

Author 6: Hafizha Ramadayanti

Author 7: Jessica Almalia

Author 8: Muhammad Putra Harifin

Keywords: EfficientNet-B0; explainable AI; Grad-CAM; transfer learning; waste classification

Paper 56: RoBERTa-Enhanced Actor–Critic Reinforcement Learning for Adaptive and Personalized ESL Instruction

Abstract: Technology-Assisted Language Learning (TALL) has developed and has greatly transformed the way English as a Second Language (ESL) is taught. The current digital resources and smart solutions have enabled more interactive and accessible learning, providing learners with an opportunity to train their skills at any time and place. Nevertheless, most of the current systems remain based on strict rules or conventional supervised training approaches. These methods can demand large quantities of labelled data, are inflexible in the learning process, and have little in the way of individualized feedback. Consequently, students may remain inattentive, and the acquisition of all the necessary language skills, such as reading, writing, listening, and speaking, may be unequal. In order to address such shortcomings, this study presents T-RLNN (RoBERTa-based Reinforcement Learning Neural Network), which is a dynamic model of ESL teaching. T-RNN combines deep contextual language comprehension and reinforcement learning in order to customize teaching to every learner. The RoBERTa encoder can retrieve semantic and syntactic feedback on responses of learners, and an actor-critic reinforcement learning agent can modify teaching plans in real time. The agent takes into account the learner-specific factors, i.e., proficiency, response time, engagement, and interaction behavior, to give the best guidance. It was trained in Python using PyTorch and tested on a curated dataset of 5,000 responses of a learner in reading, writing, listening, and speaking tasks. T-RLNN performed better than conventional models, such as Support Vector Machines, random forests, and conventional deep neural networks, with a 94.8 % accuracy, 92.7 % F1 -score, and 71.5 % Adaptivity Index. These findings indicate that T-RLNN has the potential to provide insightful, interactive, and learner-oriented ESL training and open the way to smarter and more adaptable language learning systems.

Author 1: Angalakuduru Aravind

Author 2: A. Swathi

Author 3: Jillellamoodi Naga Madhuri

Author 4: R. Aroul Canessane

Author 5: K. Lalitha Vanisree

Author 6: Elangovan Muniyandy

Author 7: Rasha M. Abd El-Aziz

Keywords: English as a Second Language; adaptive learning; reinforcement learning; RoBERTa; intelligent tutoring systems

Paper 57: Exploring Cyber Trends and Threats Towards V2X Connected Vehicles in Malaysia: A Systematic Literature Review

Abstract: The rapid expansion of 5G enabled Vehicle to Everything (V2X) communication has evolved into an intelligent transportation system by supporting applications such as autonomous driving, real-time traffic optimization, and road safety management. However, the growing connectivity and diverse communication protocols also create major cybersecurity challenges, especially in the network tier of connected vehicles. This study conducts a systematic literature review following the PRISMA framework to examine cybersecurity threats and detection models in Malaysia's V2X ecosystem. It involves an analyzing phase towards 85 peer-reviewed studies published between 2016 and 2025. This addresses three research questions: (RQ1) What is the state-of-the-art in CVs in the aspect of network technology in Malaysia, (RQ2) What are the cybersecurity trends and threats towards CVs in the network tier, and (RQ3) What are the existing models in detecting and responding to cyber threats against CVs? Study identifies critical threats, including spoofing, jamming, and denial of service attacks, while evaluating intrusion detection systems that use machine learning, deep learning, and hybrid approaches. The existing approaches are yet to face limitations in real-time performance, contextual accuracy, and supply chain resilience under Malaysia's tropical urban conditions. This study proposes a conceptual model, the SCARF-V2X model, an NGSOC integrated concept that utilizes SIEM, SOAR, and Malaysian cyber threat intelligence platforms to enable automated detection and first-layer auto-response, specifically towards supply chain threats in CVs. The proposed model aims to improve Malaysia's V2X cybersecurity landscape and introduces a proactive and adaptive model to protect CVs against evolving cyber threats.

Author 1: A’in Hazwani Ahmad Rizal

Author 2: Noor Afiza Mat Razali

Author 3: Sakinah Ali Pitchay

Author 4: Taqiyuddin Anas

Keywords: 5G; V2X; connected vehicles; cybersecurity; intrusion detection systems; anomaly detection; spoofing; DoS; network tier; machine learning

Paper 58: Benchmarking Large Language Models for Dental Clinical Decision Support: A BERT Score Analysis of Claude Opus 4.5

Abstract: The integration of Large Language Models (LLMs) into clinical decision support systems represents a significant advancement in healthcare informatics. This study presents a comprehensive evaluation framework for benchmarking LLM-generated dental treatment recommendations using BERT Score as the primary semantic similarity metric. We evaluated Claude Opus 4.5 as a Clinical Decision Support System (CDSS) across 116 dental case reports extracted from the Case Reports in Dentistry journal (2024-2025), spanning nine dental specialties. The BERT Score was calculated using the RoBERTa-large model to measure semantic alignment between AI-generated treatment plans and gold-standard published treatments. Results demonstrated strong semantic alignment with a mean BERT Score F1 of 0.8199 with a standard deviation of 0.0144 (95 per cent confidence interval: 0.8172-0.8225), significantly exceeding the 0.80 threshold (t = 14.90, p < 0.001, d = 1.38). Cross-specialty analysis revealed consistent performance across all nine dental domains (Kruskal-Wallis H = 3.07, p = 0.879), indicating robust generalizability. A significant negative correlation was observed between BERT Score and response time (ρ = -0.371, p < 0.001), suggesting a speed-accuracy trade-off in LLM reasoning. This study contributes a reproducible benchmarking methodology for evaluating LLM performance in specialized clinical domains and demonstrates the potential of BERT Score as a scalable evaluation metric for AI-generated clinical text.

Author 1: Achmad Zam Zam Aghasy

Author 2: Muhammad Lutfan Lazuardi

Author 3: Hari Kusnanto Josef

Keywords: BERT Score; Large Language Models; clinical decision support system; semantic similarity; Claude Opus 4.5

Paper 59: Understanding IT Product Purchasing Behavior of MSMEs Using Sequential Pattern Mining Approaches

Abstract: Sequential pattern mining is a crucial analytical method for understanding purchasing behavior and uncovering hidden patterns in transactional data. Unlike most prior studies that apply sequential pattern mining primarily in consumer-oriented retail settings or evaluate algorithms in isolation, this study investigates IT product purchasing behavior among Small and Medium Enterprises (SMEs) within a B2B digital transformation context through a direct comparative evaluation of three widely used algorithms: Apriori, PrefixSpan, and CloSpan. A series of controlled experiments was conducted on the same transactional datasets to assess algorithm performance in terms of accuracy, computational efficiency, and redundancy reduction. The results show that Apriori discovers exhaustive patterns at the cost of higher computational complexity, PrefixSpan achieves faster sequence extraction with balanced accuracy, and CloSpan effectively reduces redundancy by generating closed sequential patterns. Beyond pattern discovery, this study translates support, confidence, and lift metrics into actionable decision-support insights, highlighting how different algorithmic characteristics can be aligned with retention strategies, service bundling, and targeted interventions. These findings provide distinct methodological and practical contributions by positioning sequential pattern mining as a data-driven decision-support tool to accelerate digital transformation initiatives among SMEs in the IT product ecosystem.

Author 1: Rendro Kasworo

Author 2: R. Rizal Isnanto

Author 3: Budi Warsito

Keywords: Sequential pattern mining; Apriori; PrefixSpan; CloSpan; SMEs; digital transformation

Paper 60: Collaborative Dual-Framework Defense: CTI and LLM-Based Enhanced Smishing Detection

Abstract: Smishing has become a severe cybersecurity threat. Attackers now use AI and social engineering to craft more sophisticated campaigns. To address this challenge, this study proposes a dual-layer detection framework. It combines cyber threat intelligence (CTI), machine learning, and a large language model (LLM). The framework uses 22 features built from 2,811 real SMS messages. These features are categorized as content-based, context-based, and Indicators of Compromise (IOC)-based features. Five machine learning models were evaluated. XGBoost, trained with a 70% training, 10% validation, and 20% test split, achieved the best performance. It had a recall of 92.08% and an F1-score of 94.66%. For borderline cases, the study experimented with 4 LLMs (including GPT-4o and LLaMA 3). They served as a semantic verification layer. All models achieved a recall rate above 98.5% and produced human-readable explanations. The study demonstrated that these 4 models are complementary verifiers rather than main classifiers. The results show that structured threat intelligence used during feature engineering improves machine learning model performance. With semantic reasoning, the framework also generates accessible reports for non-specialists. This lowers the barrier for effective smishing detection.

Author 1: Li Guangliang

Author 2: Kalaivani Selvaraj

Author 3: Mahinderjit Singh

Keywords: Smishing detection; cyber threat intelligence; XGBoost; semantic verification; large language model

Paper 61: Development of Image Processing Filters for Improving Visibility of Fine Dentoalveolar Structures in Dental Cone-Beam Computed Tomography Images

Abstract: Cone-beam computed tomography (CBCT) imaging in dentistry requires post-reconstruction image processing to enhance diagnostic quality while minimizing radiation exposure. Visualization of fine dentomaxillofacial structures, particularly the inferior alveolar canal (IAC) and dental pulp canals, presents significant diagnostic challenges in low-dose CBCT imaging. This study investigates the application of Wiener and adaptive Wiener (LLMMSE) filters in the reconstruction domain to improve the visibility of these critical anatomical structures in low-dose dental CBCT images. Two CBCT examinations of a dry mandible were acquired using reference and low-dose protocols. The low-dose post-reconstruction data was processed using six different filters: geometric mean, LLMMSE with additive noise 15, LLMMSE with additive noise 5, moving average, Wiener, and local contrast filters. These computationally efficient filters offer practical advantages over existing complex and costly noise reduction schemes. Subjective evaluation by an experienced oral and maxillofacial radiologist demonstrated that the IAC was clearly identifiable in all low-dose datasets regardless of filter application. However, the highest visibility of narrow pulp canals was achieved with the Wiener and LLMMSE_5 filters. This proof-of-concept study demonstrates the potential of Wiener and LLMMSE_5 techniques for improving visibility of narrow dental pulp canals in low-dose CBCT images, which has important implications for endodontic diagnosis and treatment planning while supporting radiation dose reduction strategies.

Author 1: Muhannad Almutiry

Author 2: Asma’a Al-Ekrish

Author 3: Saleh Alshebeili

Keywords: Cone-beam computed tomography; CBCT; image processing; Wiener filter; LLMMSE filter; dental imaging; noise reduction; low-dose imaging; endodontics; pulp canal visualization

Paper 62: Blockchain-Based Audit Trails: Improving Transparency and Fraud Detection in Digital Accounting Systems

Abstract: Blockchain technology has emerged as a transformative innovation in digital accounting, offering robust mechanisms to enhance auditability, data integrity, and fraud prevention. This study examines how blockchain-based audit trails can improve transparency and strengthen fraud detection within modern accounting information systems. Adopting a conceptual–analytical research design supported by secondary empirical evidence, the study analyzes data drawn from recent peer-reviewed case studies, industry reports, and documented implementations of permissioned blockchain systems in auditing and financial reporting contexts. The analysis focuses on core blockchain characteristics—immutability, decentralization, cryptographic security, real-time verification, and transaction mining—and evaluates their implications for audit processes and governance mechanisms. The results highlight that blockchain-enabled audit trails allow continuous access to verified transactional data, significantly improving early detection of anomalies, reducing opportunities for data manipulation, and enhancing the reliability of financial reporting. The study further demonstrates that permissioned blockchain architectures, combined with smart contract automation, can operationally support real-time audit logging and procedural compliance while minimizing human error. However, empirical insights also reveal critical implementation challenges, including interoperability constraints, scalability issues, regulatory uncertainty, and organizational resistance. In terms of contribution, this research offers a conceptual and methodological contribution by developing an integrated blockchain-based auditing framework that systematically links technological features with audit objectives and fraud prevention mechanisms. Unlike prior descriptive reviews, this study explicitly positions its framework against existing auditing and blockchain literature, clarifying how blockchain-based audit trails extend current auditing theory and provide practical design implications for enterprise accounting systems. Overall, the findings advance scholarly understanding of blockchain-enabled auditing and provide actionable insights for auditors, system designers, and regulators seeking to implement next-generation digital audit infrastructures.

Author 1: Neni Maryani

Author 2: Munawar Muchlish

Author 3: Roza Mulyadi

Author 4: Nurhayati Solehah

Keywords: Blockchain; audit trails; fraud detection; digital accounting; transparency; data mining

Paper 63: Optimizing the Accuracy of Alzheimer's Detection Using Machine Learning and Intelligent Feature Selection Strategies

Abstract: Alzheimer’s disease is a progressive neurodegenerative disorder for which early detection remains a significant challenge due to the complexity of clinical features and the high dimensionality of medical data. This study aims to improve the accuracy and reliability of Alzheimer’s disease detection by evaluating the performance of multiple machine learning algorithms integrated with intelligent feature selection strategies. Five classification models, Decision Tree, Naïve Bayes, Random Forest, Logistic Regression, and Deep Learning, were investigated under two experimental scenarios: without feature selection and with feature selection using Recursive Feature Elimination, Binary Particle Swarm Optimization, and Variance Threshold. Model performance was evaluated using K-fold cross-validation based on accuracy, precision, recall, and F1-score metrics. The results demonstrate that feature selection consistently enhances classification performance, particularly for conventional machine learning models such as Random Forest and Logistic Regression. Although the Deep Learning model achieves competitive accuracy, its reduced precision and F1-score indicate limitations when applied to reduced feature spaces. These findings highlight the importance of incorporating appropriate feature selection techniques to address data complexity and improve the effectiveness of early Alzheimer’s disease detection.

Author 1: Suci Mutiara

Author 2: Siti Nur Laila

Author 3: Deppi Linda

Author 4: Sri Lestari

Author 5: Jean Antoni

Author 6: Christian Petrus Silalahi

Keywords: Machine learning; Alzheimer; Random Forest (RF); logistic regression; deep learning

Paper 64: Natural Human-Machine Interaction Using Static Hand Gestures for a Gestural Calculator System with DNN

Abstract: Hand gesture recognition (HGR) represents a real challenge for natural human-computer interaction, which aims to revolutionize the naturalness of traditional interfaces, allowing intuitive control of various devices without using a keyboard or mouse. Despite the availability of frameworks such as MediaPipe, which enable better detection and tracking, the major challenge remains interpreting gestures made with both hands in a natural operational setting. In this regard, this study presents a real-time gesture calculator that combines gestures made with both hands (see using one hand) and aims to address the problem of interpretation in arithmetic operations. By leveraging MediaPipe to classify the 21 hand landmarks, an optimized dense neural network (DNN) was developed capable of recognizing 13 distinct static gestures. The latter includes six gestures for each hand (ranging from 0 to 5) to represent all digits from 0 to 9, five mathematical symbols, and two specialized commands designed explicitly for control management. Even with a standard webcam, this model achieved 91% accuracy on a reduced dataset of gestures from both hands. Beyond gesture recognition, this work demonstrates how these gestures can be integrated into a fluid sequence for arithmetic operations.

Author 1: H. Abdelmoumene

Author 2: L. Meddeber

Author 3: O. Ghali

Author 4: Y. Amellal

Keywords: Hand gesture recognition; gestural calculator interface; DNN; MediaPipe; natural human-computer interaction

Paper 65: Community-Aware Influence Maximization for Suppressing Cryptocurrency Scam Misinformation

Abstract: Cryptocurrency fraud campaigns often rely on large-scale social-media diffusion to recruit victims, normalize false claims, and coordinate multi-level marketing behavior. This study examines the dynamics of the One Coin scam. It proposes an influence-maximization (IM)-driven workflow for identifying high-impact accounts whose intervention can reduce future misinformation diffusion. A directed Twitter engagement network from retweet/reply interactions is constructed and studied, and the accounts that should be prioritized for intervention to reduce the reach of future scam-promoting misinformation are identified. We evaluate six seed selection strategies: Degree, Betweenness, PageRank, k-core, CELF (lazy greedy), and Reverse Influence Sampling (RIS) under the classical Independent Cascade (IC) and Linear Threshold (LT) diffusion models using a weighted-cascade parameterization when ground-truth transmission probabilities are unavailable. Across the tested seed budgets, CELF achieves the highest expected spread, but with the highest computational cost. At the largest seed budget, Degree is effectively tied with CELF (within 0.09% under LT and 1.4% under IC), indicating a hub-dominated engagement structure in which simple reach-based heuristics can be highly competitive. RIS provides a strong quality–efficiency trade-off, remaining within approximately 9.7% (LT) and 9.5% (IC) of CELF while requiring substantially less computation. We further introduce a community-aware variant using Leiden partitions and proportional seed allocation to improve cross-community coverage; at larger budgets, this improves methods sensitive to seed over-concentration, increasing LT spread by about 9.8% for k-core and 8.6% for RIS. Overall, the results quantify practical trade-offs between spread and runtime for deployable suppression workflows and show when community-aware planning better aligns with the heterogeneous structure of scam recruitment ecosystems.

Author 1: Naglaa Mostafa

Author 2: Hatem Abdelkader

Author 3: Asmaa H.Ali

Keywords: Influence maximization; misinformation suppression; cryptocurrency scams; OneCoin; diffusion models; Leiden; community-aware seeding

Paper 66: Enhancing Business Cybersecurity Through Integrated Defense and Incident Response: A Comparative Decision Framework

Abstract: Business operations increasingly depend on digital workflows, hybrid infrastructures, and third-party ecosystems, making cybersecurity incidents a direct business continuity and governance problem rather than solely a technical concern. This paper proposes an integrated cyber defense and defense-to-response decision framework for organizations seeking to reduce exposure to external attacks and unauthorized access while improving incident detection, containment, and recovery. The framework aligns governance and control selection with NIST Cybersecurity Framework (CSF) 2.0, operational incident response considerations with NIST SP 800-61 Revision 3, control requirements with ISO/IEC 27001:2022, prioritized safeguards with CIS Controls v8.1, and adversary-behavior mapping with the MITRE ATT&CK Enterprise Matrix. We define an evaluation model that combines 1) coverage mapping across prevent-detect-respond-recover functions, 2) multi-criteria decision analysis (MCDA) for cost, complexity, and risk reduction trade-offs, and 3) a playbook-oriented response design for high-frequency attack paths relevant to business environments. A worked comparative example demonstrates how three strategy bundles (traditional perimeter controls, defense-in-depth with SIEM, and a Zero Trust + EDR + SOAR approach) can be ranked using weighted criteria and incident lifecycle metrics. The paper concludes with an implementation roadmap and measurement plan to convert the framework into an evidence-based program that supports executive decision-making and continuous improvement.

Author 1: Jurgen Mecaj

Keywords: Cyber defense; incident response; business continuity; NIST CSF 2.0; SP 800-61r3; ISO/IEC 27001:2022; CIS Controls v8.1; MITRE ATT&CK; Zero Trust; EDR; SOAR; MCDA

Paper 67: An Efficient Skin Cancer Stage Diagnostic Approach Using Customized Inception V3 Deep Learning Model

Abstract: Among all stages of skin cancer, Melanoma, Basel Cell Carcinoma (BCC), and Squamous Cell Carcinoma (SCC) have a significant impact on world health. Although deep learning offers promising potential for dermatological categorization, only limited disease groups have benefited, since most studies focus on particular illnesses rather than covering comprehensive human skin problems. Computerized analysis has been used in the past to identify cancer in skin lesion images, but challenges still persist mainly due to the multiple forms, textures, and sizes of lesions that complicate skin cancer classification. This research paper presents a Convolutional Neural Network (CNN) model customized to meet our requirements by using a pre trained InceptionV3 model along with Bayesian hyperparameter tuning. Using the ISIC 2024 and HAM 10000 datasets, the main objective is to classify skin lesions and differentiate between malignant Melanoma, BCC, and SCC. By implementing this customized model, the issue caused by variations in lesion appearance is effectively addressed, leading to more accurate predictions. Using Bayesian hyperparameter tuning can increase identification while decreasing computational cost. The proposed model performed strongly on the combined datasets by achieving combined average accuracy of 95.1 %, a precision of 94.42 %, a sensitivity of 97.3 %, a specificity of 98.8 %, and an F1 score of 95.7 %. These results demonstrate that the model significantly outperformed existing techniques and provided more accurate and consistent diagnosis of pigmented skin lesions compared to current standards.

Author 1: Adnan Afroz

Author 2: Shaheena Noor

Author 3: Muhammad Umar Khan

Author 4: Shakil Ahmed Bashir

Keywords: Melanoma; Basal Cell Carcinoma (BCC); Squamous Cell Carcinoma (SCC); Inceptionv3; Bayesian tuning; skin cancer classification

Paper 68: Exploring the Impact of Gamified Artificial Intelligence–Driven English Vocabulary Learning Systems on Learner Retention and Motivation

Abstract: The growing development of digital learning platforms has posed an increased demand on gamified and artificial intelligence-based methods of enhancing English vocabulary learning. However, existing studies often treat gamification and AI as loosely pair components, relying on static game mechanics or post-hoc analytics that limit personalization, adaptability, and long-term learning impact. To address these limitations, this study proposes the Gamified AI-Driven Vocabulary Retention and Motivation Enhancer (GAI-VRME), an adaptive learning framework that integrates machine-learning–based learner modeling, real-time difficulty calibration, and adaptive gamification strategies. In contrast to the previous systems, GAI-VRME can dynamically regulate the complexity of the task, the frequency of feedback and the sequence of rewards according to the performance and the motivational state of a specific learner, and can thus be constantly customized to the individual level as the process of learning progresses. The implementation and empirical assessment of the framework were conducted with the help of Python, TensorFlow, and Jupyter Notebook and Teaching-Learning Gamification Dataset of Mendeley Data. Mixed method analysis of vocabulary retention with paired t-tests and sentiment-analysis-based motivation modelling was used. The experimental outcomes show that GAI-VRME has much higher predictive accuracy, vocabulary retention, and learner motivation than the traditional gamified systems. These findings provide empirical evidence that deeply integrated AI-driven adaptive gamification, jointly optimizing cognitive retention and affective engagement, offers a scalable and pedagogically robust solution for modern digital vocabulary learning environments.

Author 1: Gogineni Aswini

Author 2: Madhu Munagala

Author 3: V. Saranya

Author 4: Keerthana R

Author 5: Aseel Smerat

Author 6: Vinisha Sumra

Author 7: Ahmed I. Taloba

Keywords: AI-driven gamified learning; adaptive educational systems; English vocabulary acquisition; learner motivation and engagement; vocabulary retention

Paper 69: Enhancing Misinformation Detection on Twitter with a Content-Based Multi-Lingual Bert Model

Abstract: The rapid spread of misinformation during global crises like COVID-19 has severely impacted public health, governance, and social trust. Social media platforms such as Twitter have amplified this issue, underscoring the urgent need for multilingual, real-time misinformation detection. The proposed Content-based Attention Multi-lingual BERT (CA-BERT) model addresses this challenge by enhancing the standard BERT framework with a content-based attention mechanism that assigns adaptive weights to semantically important tokens often linked to false or misleading content. This attention enables deeper contextual understanding of misinformation cues across diverse linguistic contexts. Using the LIME interpretability method, CA-BERT provides transparent explanations of its predictions, supporting accountable decision-making for policymakers and content moderators. Leveraging multilingual BERT (mBERT) allows the model to handle multiple languages simultaneously, ensuring robust cross-lingual applicability. Evaluations using a balanced multilingual tweet dataset on COVID-19 topics demonstrate that CA-BERT outperforms baseline models such as RoBERTa, DANN, and HANN, achieving 96% recall for true information and 95% for misinformation in English, with F1 Scores of 93% and 92%, respectively. The model maintains strong cross-lingual generalization, especially for Dutch (75% F1) and Spanish (72% F1), with slightly lower performance for Arabic due to tokenization and dialectal complexity. These results highlight CA-BERT’s adaptability while underscoring the need for improved handling of low-resource, morphologically rich languages. Future work involves region-specific preprocessing, cross-lingual transfer learning, and multimodal misinformation detection, aiming to transform CA-BERT into a core component of multilingual real-time disinformation monitoring systems.

Author 1: Krishna Kumar

Author 2: Akila Venkatesan

Keywords: Component; misinformation detection; multi-lingual BERT; content-based attention mechanism; syntactic-semantic similarity; explainable AI; LIME interpretability; COVID-19 misinformation; cross-lingual generalization; twitter; adversarial robustness

Paper 70: MQTT Broker Congestion Mitigation Using Huffman Deep Compression

Abstract: This study presents an improved MQTT protocol designed to address broker congestion and connection overflow in large-scale IoT networks. The proposed method integrates Huffman Deep Compression (HDC) at the publisher side to mitigate network traffic and latency. Unlike standard MQTT, which suffers from broker overload, our approach applies efficient data compression on resource-constrained sensor devices prior to publishing. The proposed approach was validated on a real-world air pollution dataset collected from the Tanjung Malim monitoring station in Malaysia, using ESP8266-based IoT nodes. Experimental results demonstrated that broker congestion was reduced by 84.26% for QoS 0 and 79.6% for QoS 1, significantly outperforming both standard MQTT and the state-of-the-art MRT-MQTT (58% and 45%, respectively). The method attained a high compression ratio of 2.62, which directly led to a dramatic reduction in power consumption from 2,664,864 to 63,216 mA (QoS 0) and from 3,155,760 to 49,168 mA (QoS 1). This substantial saving in current consumption contributes to extended device lifetime and enhanced energy efficiency. The findings highlight the potential of this enhanced protocol to support massive IoT deployments by minimizing network overhead at the broker.

Author 1: Ammar Nasif

Author 2: Zulaiha Ali Othman

Author 3: Nor Samsiah Sani

Author 4: Yousra Abudaqqa

Keywords: Compression; network congestion; connection overflow; deep learning; IoT; broker congestion; IoT network; sensor; latency reduction; publishers; broker; MQTT; power consumption

Paper 71: Advancing Blood Supply Chain Prediction Based on a Novel Hybrid Machine Learning

Abstract: Blood supply chains constitute a critical yet often overlooked component of modern public health systems, as they coordinate donors, collection centers, hospitals, and patients. One of the major operational challenges lies in planning the deployment of mobile blood collection units under highly variable and uncertain spatio-temporal demand. In this context, this study proposes a novel hybrid machine learning framework for predicting donor return potential and supporting location and time selection decisions for mobile blood drives. The proposed approach combines Support Vector Regression (SVR) and Light Gradient Boosting Machine (LGBM) through a dynamic, context-aware weighting function designed to capture both temporal regularities and nonlinear spatial heterogeneity in donor behavior. The model is evaluated using real-world data collected from a blood collection center operating multiple mobile units. Experimental results demonstrate that the proposed hybrid framework consistently outperforms its individual components, achieving R² values of up to 83% for certain locations, together with low Mean Absolute Error (MAE) and Mean Squared Error (MSE). These results confirm the robustness and stability of the proposed approach. Beyond predictive performance, the model is intended to be integrated into a decision-support system to help managers optimize logistical resources and improve the strategic planning of mobile blood collection campaigns. This work contributes to the emerging field of data-driven blood supply chain optimization by introducing a spatio-temporal, hybrid predictive core specifically designed for operational decision support.

Author 1: Chaimae Mouncif

Author 2: Mohamed Amine Ben RABIA

Author 3: Adil Bellabdaoui

Keywords: Blood supply chain; mobile blood collection units; spatio-temporal prediction; hybrid machine learning; decision support

Paper 72: An Integrated Carbon Footprint Calculation System Model for Net Zero Emission in the Manufacturing Industry Based on GHG Protocol and DEFRA

Abstract: Manufacturing industries play a critical role in achieving Net Zero emission targets due to their significant contribution to greenhouse gas emissions. However, existing carbon footprint calculation practices often apply the GHG Protocol and emission factor standards independently, resulting in fragmented methodologies and limited decision-support capabilities. This study develops a carbon footprint calculation system model that integrates GHG Protocol emission scope classification with DEFRA emission conversion factors, supported by a decision-support framework for Net Zero emission planning. Using a Design Science Research (DSR) approach, the study produces a conceptual system model that structures activity data, emission scope classification, and standardized carbon calculation logic into a unified framework. The proposed model enables transparent aggregation of emissions across Scope 1, Scope 2, and Scope 3, while the decision-support framework translates calculation results into decision variables, scenario-based analysis, and Net Zero roadmap formulation. The system functions as a decision-support system that assists manufacturing organizations in interpreting carbon footprint results and supports Net Zero emission planning. The findings demonstrate that integrating standardized carbon accounting methodologies within a system-oriented design enhances methodological coherence, traceability, and strategic relevance for sustainability decision-making in the manufacturing sector.

Author 1: Dinar Rahayu

Keywords: Carbon footprint; net zero emission; GHG protocol; DEFRA; decision support system

Paper 73: AI-Based Parkinson’s Disease Diagnosis and Prediction with Therapeutic Game Design for Engagement

Abstract: Parkinson’s disease (PD) is a progressive neurodegenerative disease that impacts motor and cognitive functions, and early diagnosis and management are essential to enhance patient outcomes. The study assumes the implementation of Artificial Intelligence (AI)-based diagnostic and predictive algorithms, along with therapeutic game design, to assist patients in improving the management and treatment of PD. The existing approaches to PD diagnosis rely heavily on clinical observation of symptoms and on traditional imaging methods, which may be subjective, time-consuming, and prone to human error. Moreover, conventional interventions are not consistently engaging or tailored to the patient, and hence, treatment adherence is not optimal. To overcome these difficulties, we present PD in the framework of AI (PD-AI), leveraging machine learning algorithms to enhance early diagnosis and predict disease progression. The system will be implemented as a mobile app that integrates AI with therapeutic gaming, with real-time symptom tracking based on sensor readings (e.g., tremors, motor skills) and interactive therapeutic games provided to the patient to maintain their engagement. The suggested approach enhances early diagnosis rates, provides a tailored approach, supports continuous monitoring of symptoms, and encourages patients to follow their treatment actively. An active, efficient, and convenient management strategy is facilitated by data analysis based on frequent examinations and feedback via the app. Preliminary results indicate that the PD-AI model improves case diagnosis accuracy and patient compliance with treatment regimens, demonstrating its effectiveness for both medical experts and patients with PD.

Author 1: Marwah Muwafaq Almozani

Author 2: Hüseyin Demirel

Keywords: Parkinson’s disease; Artificial Intelligence (AI); mobile application; early diagnosis; machine learning algorithms

Paper 74: A Systematic Review and Taxonomy of Privacy-Preserving Blockchain Consensus Mechanisms

Abstract: Blockchain systems rely on consensus mechanisms to validate transactions and coordinate distributed participants, making consensus a critical layer that shapes security, trust, and privacy. Although blockchain is increasingly applied in privacy-sensitive domains such as healthcare, smart cities, and the Internet of Things, existing review studies primarily examine security or performance and rarely analyse how consensus-level design properties influence privacy risks. As a result, privacy is often treated as a peripheral enhancement rather than a core consensus concern. This study presents a systematic literature review that examines blockchain consensus mechanisms from a privacy-focused perspective. The review aims to identify which consensus classes are most commonly used in privacy-preserving blockchain systems, what privacy limitations are reported across different consensus designs, and how privacy-preserving techniques are integrated into consensus mechanisms. The review follows PRISMA and Kitchenham-guided procedures, using structured search and screening of peer-reviewed journal articles from major academic databases, followed by relevance and quality assessment. 72 peer-reviewed journal articles were synthesised using taxonomy-based and thematic analysis. The proposed taxonomy explicitly classifies studies by consensus mechanism class, privacy limitation, and integration level, enabling structured comparison beyond existing surveys. The findings show that Byzantine Fault Tolerant (BFT)-based consensus mechanisms are most frequently adopted in privacy-preserving blockchain applications. However, privacy challenges such as identity exposure and communication pattern leakage remain common and are closely linked to consensus design properties. In addition, most studies rely on external privacy mechanisms rather than embedding privacy directly into the consensus layer. This review contributes a structured taxonomy, clear analytical insights, and practical guidance that support the development and evaluation of privacy-aware blockchain consensus mechanisms.

Author 1: Sahnius Usman

Author 2: Sharifah Khairun Nisa Habib Elias

Author 3: Suriayati Chuprat

Author 4: Ahmad Akmaluddin Bin Mazlan

Keywords: Blockchain; Byzantine Fault Tolerance; consensus mechanisms; privacy-aware consensus; privacy preserving; systematic literature review

Paper 75: Experimental Validation of Contextual Parameters and Comparative Analysis with State-of-the-Art in CARS Recommendation Systems in Ubiquitous Computing

Abstract: The most important role of Consumer Behavior prediction plays in e-commerce, various ways of marketing, and Context-aware Recommendation Systems (CARS). From an Amazon consumer dataset, we conduct a comparative analysis of different machine learning models to compare their performance or effectiveness in predicting consumer behavior based on an Amazon consumer dataset. Additionally, we introduce a new algorithm combining feature selection and optimization that aims to enhance prediction accuracy. Person behavior prediction has historically helped enhance e-commerce, marketing, and Context-Aware Recommendation Systems-CARS, allowing businesses to get closer to customers and understand their needs better from the time they appeared to the time an analysis could be done. The research work performs comparative analysis among various machine learning techniques, like Logistic Regression, Decision Tree, Random Forest, SVM, and KNN, to see which one is more effective in predicting customer behavior, based on an Amazon consumer dataset. Besides, a new algorithm that merges feature selection and optimization is proposed and implemented to guarantee better prediction accuracy. The project is aimed at the creation of data-driven decision systems powered by an optimized machine learning framework for customer analytics.

Author 1: Pranali Gajanan Chavhan

Author 2: Ritesh Vamanrao Patil

Keywords: Context-aware recommendation systems; multi-modal recommendation; transformer-based models; experimental validation; contextual parameter modeling; user behavior modeling

Paper 76: Edge Artificial Intelligence for Real-Time Fresh Produce Identification in Retail Weighing Systems

Abstract: Real-time recognition of loose fresh produce is a key requirement for intelligent retail weighing systems, enabling automated replacement of or assistance to manual PLU-based item selection. However, the deployment performance of recent YOLO architectures on embedded edge platforms such as the NVIDIA Jetson Xavier NX remains insufficiently studied in practical retail scenarios. This study aims to benchmark recent YOLO architectures for real-time fresh produce recognition on embedded edge devices. This work presents an Edge–AI retail weighing system that recognizes Malaysian fresh produce using YOLOv9, YOLOv10, and YOLOv11 models on the Jetson Xavier NX. A domain-specific dataset of 8 450 images across 26 classes was created by merging ImageNet and Roboflow sources and applying quality filtering and unified preprocessing. Each model was fine-tuned and optimized with TensorRT at FP16 and INT8 precision. Transfer learning improved accuracy across all models; YOLOv11-Large achieved the highest mAP@0.5 of ≈ 0.897 but at a reduced frame rate, while the mid-sized YOLOv10-M delivered an mAP@0.5 of ≈ 0.890 with near-real-time performance inference. Inference analysis shows that pre-and post-processing add only a few milliseconds per frame yet become proportionally significant as inference speeds increase; YOLOv11’s Non-Maximum Suppression (NMS) head introduces notable latency relative to YOLOv10’s NMS-free design. Quantized YOLOv10-M and YOLOv10-N sustain ≈ 14–19 FPS , offering the best balance between accuracy and speed. Qualitative tests on market footage confirm robust detection, indicating that these optimized models enable accurate, low-latency produce identification for intelligent retail weighing.

Author 1: Shi Han Teo

Author 2: Jun Kit Chaw

Keywords: Real-Time object detection; Edge Artificial Intelligence (Edge AI); transfer learning; YOLO algorithm; nvidia jetson; fresh produce recognition; TensorRT optimization; model quantization

Paper 77: Calibrated Residual Intelligence for Intra-Procedural CBCT–Based Collateral Grading in Ischemic Stroke

Abstract: Brain stroke occurs when the brain’s blood supply is disrupted, leading to oxygen deprivation and rapid neuronal death. Ischemic stroke, the focus of this study, accounts for most cases and is strongly influenced by collateral circulation, a network of alternative vessels that stabilize perfusion when a primary artery is obstructed. Collateral status determines the extent of salvageable tissue and is typically graded manually using modalities such as magnetic resonance angiography (MRA), computed tomography (CT), and cone-beam computed tomography (CBCT), a process prone to subjectivity and inter-observer variability. This study proposes a ResNet-18–based deep learning framework for automated three-class classification of collateral circulation (Good, Moderate, Poor) from intra-procedural CBCT scans. A curated dataset of 45 patient cases (22,861 DICOM slices), annotated by an expert neuroradiologist, was preprocessed with patient-wise partitioning, normalization, and augmentation. The model achieved a validation accuracy of 88.8%, a micro-averaged precision–recall score of 0.947, and a macro-averaged ROC AUC of 0.958. Calibration analysis confirmed well-aligned probability estimates, while most misclassifications occurred in the Moderate class, reflecting inherent clinical ambiguity. Com-pared with prior CBCT studies using shallower architectures, the proposed framework demonstrates substantially higher accuracy, improved calibration, and enhanced robustness. These findings highlight the feasibility of ResNet-18 applied to CBCT imaging as a reliable and efficient tool to support neuroradiologists in collateral grading during hyperacute stroke management.

Author 1: Kazi Ashikur Rahman

Author 2: Nur Hasanah Ali

Author 3: Ahmad Sobri Muda

Author 4: Nur Asyiqin Amir Hamzah

Author 5: Noradzilah Ismail

Keywords: Collateral circulation; brain stroke; ischemic stroke; deep learning; ResNet-18

Paper 78: A Robust RT-DETR-Based Method for Complex Self-Service Buffet Scene Detection

Abstract: Object detection in buffet-style environments is highly challenging due to densely stacked tableware, frequent occlusions, strong illumination reflections, and substantial visual similarity across categories, all of which undermine the robustness of existing detectors. To address these issues, this paper proposes an improved real-time detection transformer–based model with a lightweight design while significantly enhancing multi-scale feature representation. First, a re-parameterized stem module is introduced to strengthen shallow texture extraction with negligible computational overhead. Second, a dynamic multi-kernel refinement module is developed to enrich directional texture modeling and cross-scale semantic aggregation. Furthermore, a heterogeneous-kernel feature pyramid network is constructed by integrating adaptive multi-scale fusion, multi-kernel fusion nodes, and a lightweight upsampling strategy to improve cross-level feature consistency and mitigate aliasing caused by conventional upsampling. Experimental results on a self-constructed buffet-scene dataset demonstrate that the proposed method improves mAP50 and mAP50:95 by 2.6% and 1.9%, respectively, while reducing parameters and GFLOPs by 42.6% and 42.3%, and increasing inference speed to 103.1 FPS. On Dota v1.0 and SkyFusion data sets, the small target detection ability has also been improved. The substantial reductions in computation and model size further confirm the effectiveness and practical value of the proposed approach for complex catering scenarios.

Author 1: Zhengwang Xu

Author 2: Hongyang Xiao

Author 3: Zhou Huang

Keywords: RT-DETR; lightweight object detection; multi-scale feature fusion; attention enhancement; buffet-scene perception

Paper 79: Predicting Students’ Cognitive Profiles Using Explainable Machine Learning

Abstract: Conventional educational strategies fail to comprehend and leverage the diversity of learners’ cognitive strengths and overlook their innate intelligence, a fundamental driver of learning. To address this gap, this study proposes a machine learning (ML) framework to predict students’ overall innate intelligence scores, independent of subject domain or exam structure, using the Learning Meta-Learning dataset, which includes data from 1,021 university students. Seven regression models, including Decision Tree, Random Forest, Extra Trees, Gradient Boosting, Extreme Gradient Boosting, LightGBM, and CatBoost, along with their ensembles have been trained and evaluated. Explainable Artificial intelligence (XAI) technique SHAP is used for important feature selection among 54 features and recursive feature elimination to further enhance model accuracy and interpretability. In comparison to the conventional method, the proposed SHAP-based ML approach is lightweight, trained with selected features, and has shown improvements in accuracy. The accuracy without XAI on CatBoost is 98.32%, whereas with XAI on CatBoost it is 98.53% using only 35 features out of 54. These findings suggest that integrating learners’ cognitive profile prediction model can aid the design of personalized educational strategies, moving beyond one-size-fits-all educational strategies.

Author 1: Sonia Corraya

Author 2: Fahmid Al Farid

Author 3: M Shamim Kaiser

Author 4: Shamim Al Mamun

Author 5: Jia Uddin

Author 6: Hezerul Abdul Karim

Keywords: Explainable AI; SHAP feature selection; machine learning; innate intelligence prediction; cognitive profiles; student diversity

Paper 80: Inter-Robots Position Estimation Using UWB Positioning Devices for Distributed Cooperative Control of Multiple Robots

Abstract: Ultra-wide band (UWB)-based positioning methods for static environments have been continuously improved; how-ever, many existing approaches rely on fixed reference nodes, and methods for directly computing relative positions among mobile units have not been sufficiently investigated. This paper presents a relative positioning approach for multi-robot systems using UWB wireless communication within a distributed cooperative control framework. In the proposed approach, multiple UWB positioning devices are arranged in regular polyhedral configurations to improve the uniformity of ranging accuracy. Robot coordinates are estimated using a nonlinear least-squares optimization method formulated from a system of simultaneous distance equations, enabling mutual relative position estimation among robots. Simulation experiments were conducted to evaluate estimation accuracy and error characteristics under different geometric configurations. Four configurations: square, tetrahedron, regular tetrahedron, and regular octahedron were considered, and their error magnitudes and axis-wise distributions were compared. The simulation results indicate that the proposed configuration achieve lower estimation errors than the other configurations evaluated. Based on these findings, experimental verification was performed, and the observed trends were consistent with the simulation results. This work provides a systematic investigation of a mutual positioning system that enables robots to estimate their positions with respect to one another without relying on fixed landmarks. Unlike existing method, our approach enables the determination of relative positions between robots based on distances measured by each robot. The proposed approach is expected to be applicable to autonomous decentralized control in multi-robot systems operating in static environments.

Author 1: Airi Kojima

Author 2: Kohei Yamagishi

Author 3: Tsuyoshi Suzuki

Keywords: Mobile robots; distributed cooperative control; mutual positioning system; relative position estimation; UWB position-ing devices; geometric structure

Paper 81: DAMCSeg: Dynamic Adaptive Multi-Modal Collaborative Semantic Segmentation

Abstract: While current semantic segmentation models excel in controlled environments, they often struggle with key challenges such as dynamic multi-modal data, small target recognition, and computational efficiency for edge deployment. Motivated by these limitations, this study explores targeted solutions and presents DAMCSeg (Dynamic Adaptive Multi-modal Collaborative Semantic Segmentation), an innovative framework that introduces advancements across feature fusion, training paradigms, and model efficiency. The core contributions of DAM-CSeg include: 1) a Dual-Stage Attention Fusion (DSAF) module that dynamically adjusts multi-branch fusion weights based on scene complexity; 2) an end-to-end joint training framework for object detection and semantic segmentation designed to minimize inter-stage error propagation; and 3) a Lightweight Multi-Modal Fusion (LMMF) module that efficiently integrates multi-source data with low computational overhead. To rigorously evaluate the proposed method’s effectiveness against these specific challenges, extensive experiments are conducted on mainstream benchmark datasets. The results demonstrate that DAMCSeg achieves high accuracy and operational efficiency, effectively addressing critical issues in dynamic scene adaptation, complex target segmentation, and edge device deployment. This provides a practical and viable solution for semantic segmentation in demanding applications such as autonomous driving and medical image analysis.

Author 1: Qirui Liao

Author 2: Zuohua Ding

Author 3: Hongyun Huang

Keywords: Dynamic Adaptive Multi-Modal Collaborative; Dual-Stage Attention Fusion (DSAF); End-to-End Detection-Segmentation Joint Framework; Lightweight Multi-Modal Fusion (LMMF)

Paper 82: Comparison of Histogram Equalization and Multi-Scale Retinex Methods for Near-Infrared Image Enhancement in Drowsiness Detection

Abstract: Computer vision-based drowsiness detection faces significant challenges in low-light conditions, particularly when using near-infrared (NIR) sensors for driver monitoring systems. Appropriate image enhancement methods are crucial to improve detection accuracy. This study systematically evaluates five enhancement methods: Histogram Equalization (HE), Adaptive Histogram Equalization (AHE), Contrast-Limited Adaptive Histogram Equalization (CLAHE), Brightness Preserving Dynamic Histogram Equalization (BPDHE), and Multi-Scale Retinex with Color Restoration (MSRCR). The evaluation was conducted on 4,272 frames from the University of Liège (ULg) Multimodality Drowsiness Database (DROZY) using four no-reference metrics: Natural Image Quality Evaluator (NIQE), Perception-based Image Quality Evaluator (PIQE), Shannon Entropy, and Lightness Order Error (LOE). Additional validation was performed by measuring the face detection rate using MediaPipe. The results show that CLAHE achieves an optimal balance with an NIQE of 4.61 (best natural quality), a detection rate of 97.9%, and an LOE of 0.058 (superior structural preservation). MSRCR produces the highest entropy (6.58) but the lowest detection rate (75.6%), indicating structural distortion in the NIR context. Statistical validation using the Wilcoxon signed-rank test and the Friedman test confirmed the significance of the findings (p < 0.05). CLAHE is recommended for NIR surveillance-based drowsiness detection systems.

Author 1: Moh Hadi Subowo

Author 2: Pulung Nurtantio Andono

Author 3: Guruh Fajar Shidik

Author 4: Heru Agus Santoso

Keywords: Image enhancement; near-infrared; drowsiness detection; histogram equalization; multi-scale retinex; CLAHE; no-reference quality metrics

Paper 83: Smart Agriculture in Morocco: An Intelligent Deep Learning Framework for Crop Disease Diagnosis

Abstract: The Moroccan agricultural sector is currently navigating a pivotal transformation driven by the “Generation Green 2020–2030” national strategy, which places a high priority on the digitalization of farming practices to bolster resilience against climate volatility and phytopathological risks. This study proposes a robust Smart Agriculture Framework engineered to automate crop disease diagnosis within mobile environments with limited resources. Unlike generic standard Deep Learning models often unsuited for local specificities, the methodology presented here is specifically tailored to Morocco’s agroecological context, targeting three strategic crops: Tomato (Souss-Massa region), Potato (Gharb plains), and Wheat (Chaouia region). A hybrid intelligent architecture is introduced that integrates a lightweight Convolutional Neural Network (CNN) with Particle Swarm Optimization (PSO-CNN) for autonomous hyperparameter tuning. The proposed framework was validated using a curated dataset of 15,000 images, rigorously augmented to reflect local field conditions, yielding a classification accuracy of 94.7%. This work effectively bridges the gap between theoretical AI architectures and practical Precision Farming, providing a rapid decision support system to minimize yield losses and align with the national objective of establishing a digitally empowered agricultural ecosystem.

Author 1: Hajar Krim

Author 2: Abdelhadi Assir

Keywords: Smart agriculture; deep learning; framework; Morocco; generation green; crop disease; PSO-CNN; precision farming

Paper 84: Evolution of Image Captioning Models: A Systematic PRISMA Review

Abstract: This article presents a systematic review of image captioning approaches conducted according to the PRISMA methodology, ensuring a rigorous, transparent, and reproducible analysis of the literature. The study traces the evolution of image captioning methods, beginning with early machine learning–based techniques that rely on handcrafted visual features, object detection, and template-based or statistical language models. While these approaches established foundational concepts, they are constrained by limited scalability and semantic expressiveness. Specific challenges include difficulty in capturing complex object relationships and inability to generate diverse descriptions for the same image. Image captioning represents a key research problem at the intersection of computer vision and natural language processing, aiming to automatically generate coherent and semantically accurate textual descriptions of visual content. Due to its multimodal nature and practical relevance, it has attracted increasing attention in artificial intelligence research. The review then examines the transition toward deep learning–based models, which have become dominant due to their improved performance. Encoder–decoder architectures are analyzed, highlighting the use of convolutional neural networks for visual representation and recurrent neural networks for caption generation. Attention-based models are discussed for their ability to focus on salient image regions, followed by reinforcement learning–based methods that directly optimize evaluation metrics and semantic-driven architectures that enhance caption relevance. Finally, recent advances based on Transformer architectures and large-scale multimodal pretraining are reviewed, along with key application domains and open challenges for future research in image captioning.

Author 1: Abdelkrim SAOUABE

Author 2: Khalid TIZRA

Author 3: Doha BANOUI

Keywords: Image captioning; vision-language models; semantic-based models; transformer models; attention mechanism; pre-trained models; GPT-based models

Paper 85: Interpretable Structural Stability Analysis for Long-Term Cognitive IoT Time-Series Data

Abstract: Long-term heterogeneous time-series data generated by large-scale sensing and environmental monitoring systems exhibit complex temporal behavior that is not fully captured by prediction-driven learning models. While most existing approaches emphasize short-term forecasting accuracy, comparatively little attention has been given to the analysis of long-term structural stability inherent in such data. In this work, we propose a lightweight, training-free analytical framework for quantifying structural stability in long-duration time-series using stability-preserving preprocessing and interpretable temporal statistics. The proposed method combines total variation regularization with rolling statistical analysis to assess the consistency of local temporal behavior relative to global characteristics over extended time horizons. Structural stability is quantified using a simple yet effective stability index that captures deviations between local and global temporal trends. The framework is evaluated using more than two decades of daily environmental observations, including temperature, relative humidity, and precipitation, obtained from the NASA POWER repository for a representative location in Assam, India. Experimental results demonstrate consistent and systematic reductions in the stability index following preprocessing across all variables, indicating improved temporal consistency without structural distortion. Additional robustness analysis across multiple temporal scales confirms that the proposed framework is insensitive to window size selection and preserves long-term structural behavior. These findings suggest that meaningful insights into temporal stability can be obtained without reliance on model training or predictive learning, making the proposed approach suitable for interpretable, resource-efficient analysis of long-term heterogeneous time-series data. Unlike conventional stability descriptors such as variance-based measures or correlation-based consistency metrics, the proposed stability index directly quantifies local-to-global deviation of temporal descriptors across multiple window scales, enabling interpretable and comparable stability assessment without requiring model training or forecasting error baselines.

Author 1: Basab Nath

Author 2: Yonis Gulzar

Keywords: Structural stability; training-free framework; total variation regularization; rolling statistics; stability index

Paper 86: Interpreting Multimodal Fake News Detection Models: An Experimental Study of Performance Factors and Modality Contributions

Abstract: The widespread dissemination of multimodal mis-information requires models that can reason across textual and visual content while remaining interpretable. However, many existing multimodal fusion approaches implicitly assume uniform modality reliability, providing limited transparency into modality contributions. This study introduces TweFuse-W, a lightweight multimodal framework for fine-grained fake-news detection that reframes multimodal fusion as a modality reliability estimation problem, rather than merely merging modalities or explicitly modeling their interactions. TweFuse-W integrates BERTweet-based textual representations with Swin Transformer visual features using a sample-conditioned, learnable weighted-sum gate operating at the modality level, producing global reliability weights without cross-attention overhead. By explicitly param-eterizing modality contributions during inference, the proposed approach provides intrinsic interpretability. Experiments on the six-class Fakeddit dataset show that TweFuse-W achieves a macro-F1 score of 0.838, outperforming simple concatenation (macro-F1 = 0.820). Analysis of the learned modality weights confirms meaningful interpretability, with textual representations dominating in Satire, Misleading, False Connection, and Imposter Content (αT = 0.57–0.62), while visual cues exert greater influence in Manipulated Content (αV = 0.51). Overall, these findings demonstrate that adaptive modality weighting enhances both predictive performance and model transparency, serving as a lightweight and interpretable complementary fusion strategy for multimodal fake-news detection.

Author 1: Noha A. Saad Eldien

Author 2: Wael H. Gomaa

Author 3: Khaled T. Wassif

Author 4: Hanaa Bayomi

Keywords: Multimodal fake news detection; modality reliability modeling; adaptive fusion; interpretable fusion; lightweight multi-modal models

Paper 87: Enhancing SCADA Security in Critical Infrastructure: A Multi-Layered Architecture Using IoT-Based Monitoring and AI-Driven Anomaly Detection

Abstract: Supervisory Control and Data Acquisition (SCADA) systems are central to the efficient operation of critical infrastructure such as energy, water, and industrial networks. However, the increased digital integration of SCADA components, especially through Internet of Things (IoT) technologies, has simultaneously broadened their exposure to cyber threats. This project presents a simulated SCADA system architecture designed to model, monitor, and secure real-time industrial telemetry using open-source platforms Node-RED and ThingsBoard. Leveraging real-world data collected from the Aventa AV-7 wind turbine in Switzerland, the project implements a multilayered architecture comprising edge, fog, and cloud layers, equipped with synchronized databases for integrity comparison and threat forensics. Artificial intelligence (AI) models are integrated into the system to perform anomaly detection using supervised, unsupervised, and deep learning (LSTM) algorithms. Cyberattacks including Distributed Denial of Service (DDoS), false data injection, and replay attacks are simulated to evaluate the system’s resilience. This report details each stage of the project from data preprocessing and system design to implementation and evaluation culminating in a set of strategic recommendations for enhancing SCADA security through AI-driven frameworks.

Author 1: Mohammad Alqahtani

Author 2: Abdulkarim Amin

Author 3: Kyounggon Kim

Author 4: Seokhee Lee

Keywords: SCADA security; industrial IoT; anomaly detection; machine learning; digital forensics; wind turbine telemetry

Paper 88: Lightweight Dual-YOLOv8 Instance-Aware Semantic Segmentation for Real-Time Autonomous Driving on Edge ARM/GPU Platforms

Abstract: Semantic segmentation is a fundamental component of autonomous driving systems, enabling accurate scene understanding and object-level perception. However, achieving precise instance-level delineation while maintaining real-time performance on resource-constrained platforms remains a significant challenge, particularly for edge deployment scenarios. This paper proposes a lightweight dual-YOLOv8 fusion framework for instance-aware semantic segmentation in autonomous driving applications. The proposed approach integrates YOLOv8n-seg and YOLOv8s-seg through a multi-scale fusion strategy that exploits their complementary feature representations to improve the segmentation of road-relevant objects, including cars, buses, trucks, and motorcycles. The framework is evaluated on the Reetiquetado de Vehiculos dataset using standard instance-level segmentation metrics. Experimental results demonstrate strong performance, achieving an overall mAP@0.5 of 92.9% and mAP@0.5:0.95 of 80.8%, while maintaining real-time inference with an average processing time of 7.9 ms per image (126 FPS) on an NVIDIA RTX 3050 GPU. Class-wise and confidence-based analyses confirm consistent segmentation accuracy across vehicle categories, highlighting the robustness of the proposed fusion strategy in handling scale variation, occlusions, and object diversity. In addition, an embedded deployment analysis provides insight into the feasibility and practical constraints of deploying the proposed framework on representative edge platforms. Overall, the proposed dual-YOLOv8 fusion framework achieves an effective balance between segmentation accuracy and computational efficiency, making it suitable for real-time autonomous driving perception on edge ARM/GPU platforms and Advanced Driver Assistance Systems (ADAS).

Author 1: Safa Teboulbi

Author 2: Seifeddine Messaoud

Author 3: Mohamed Ali Hajjaji

Author 4: Mohamed Atri

Author 5: Abdellatif Mtibaa

Keywords: Autonomous driving; instance-aware semantic seg-mentation; real-time instance segmentation; YOLOv8; dual-model fusion; edge deployment

Paper 89: Lightweight Machine Learning for Real-Time Gear Change Prediction in Autonomous Parking

Abstract: Real-time motion planning for autonomous parking on embedded advanced driver-assistance system (ADAS) platforms faces a fundamental computational bottleneck: transformer-based approaches (e.g., Motion Planning Trans-former, Diffusion-based planners) achieve strong performance but incur prohibitive computational costs unsuitable for resource-constrained automotive systems. This work proposes a lightweight alternative machine learning approach using Random Forest classifiers and regressors to predict parking trajectory regions and vehicle orientations, enabling accelerated Rapidly-exploring Random Trees (RRT) planning without sacrificing robustness. The approach is trained on a dataset of 10,725 synthetic per-pendicular backward parking scenarios generated via Rapidly-exploring Random Tree Star (RRT*) in the Reeds-Shepp con-figuration space. Using Random Forests with 20 trees and maximum depth 8, the method achieves 98.3–100% success rate in multi-direction-change scenarios with planning times of 0.15–0.25 seconds, compared to 2.81 seconds for unconstrained RRT. In scenarios with insufficient prediction guidance, the constrained planner can maintain a fallback mechanism that preserves RRT’s probabilistic completeness guarantees. This work demonstrates that simpler machine learning models can match transformer-based approaches while remaining practical for embedded deployment.

Author 1: Ahmed A. Kamel

Author 2: Reda Alkhoribi

Author 3: M. Shoman

Author 4: Mohammed A. A. Refaey

Keywords: Autonomous parking; direction change detection; embedded systems; machine learning; motion planning; random forest; rapidly-exploring random trees; rapidly-exploring random tree star

Paper 90: Relative Position Estimation for Multi-Robots Based on Vertex Distance Between Regular Tetrahedral Units

Abstract: Optimal decentralized cooperative control in multi-robot systems requires simultaneous local sensing and inter-agent communication. Ultra-wide band (UWB) wireless communication has been investigated as a self-contained positioning system capable of supporting both functions within a single device. Conventional UWB-based positioning methods estimate absolute positions using distance measurements relative to fixed anchors in the environment; moreover, while relative position estimation methods based on antenna configurations have been studied, mutual relative position estimation with respect to the reference coordinate frames of the agents themselves has not yet been investigated. To address this gap, this paper proposes a three-dimensional relative position estimation method for distributed cooperative control based on inter-vertex distances sharing. In the proposed system, inter-vertex distances between units are measured, where each unit is equipped with four UWB devices arranged in a regular tetrahedral geometric structure. An optimization-based estimation process is applied and enhanced with a k-means clustering method to mitigate convergence to local minima. The estimation accuracy was evaluated through both simulations and real-world experiments. The results demonstrate that the proposed method can accurately estimate relative positions between units and is effective for multi-robot systems operating on planar surfaces.

Author 1: Airi Kojima

Author 2: Kohei Yamagishi

Author 3: Tsuyoshi Suzuki

Keywords: Multi-robot system; distributed cooperative control; relative position estimation; self-contained positioning system; inrter-vertex distance; geometric structure

Paper 91: Explainable AI for Enhancing Awareness of Academic Stress Among International University Students

Abstract: Academic stress is a common challenge in higher education, especially for international university students who must adapt to new academic systems, expectations, and learning environments. In recent years, artificial intelligence has been increasingly used to analyze academic data and estimate student stress. However, most AI-based systems prioritize prediction accuracy over providing valuable support for student understanding. As a result, students may receive stress-related indicators without a clear explanation of how these results relate to their academic tasks or activities. This state-of-the-art review discusses current research on explainable artificial intelligence in the field of academic stress and student awareness. Based on literature published between 2020 and 2025, this review synthesizes work from educational technology, learning analytics, and explainable AI from a Human–Computer Interaction perspective. The analysis focuses on the representation of academic stress, the design of explanatory frameworks, and the extent to which existing systems facilitate students’ ability to interpret and reflect on their work. The review finds that awareness is rarely treated as an explicit outcome in existing research. Although explainable models are increasingly used, the explanations they produce are often technical and not student-oriented. International students are an underrepresented group in the literature, despite the apparent differences in their academic preparation, linguistic ability, and expectations. Consequently, these shortcomings limit the effectiveness of artificial intelligence systems as tools for enhancing student awareness. This review highlights the need to shift from prediction-oriented approaches toward awareness-oriented explainable AI systems that prioritize student understanding. By emphasizing human-centered explanation design and inclusive evaluation, future research can better support students in making sense of academic stress within diverse higher education environments.

Author 1: Ahmed Almathami

Author 2: Richard Stone

Keywords: Explainable artificial intelligence; academic stress; student awareness; international university students; learning analytics; Human–Computer Interaction

Paper 92: TRI-GATE: A Tri-Modal Anti-Spoofing System for Gate Access Using Vehicle, License Plate, and Face Recognition

Abstract: Vehicle gate access, in general, still relies heavily on manual inspection of identification cards and visual verification by security guards, which is slow, tedious, and susceptible to spoofing. Single-modality, computerized systems that utilize license plates, vehicle appearance, and facial recognition can partially alleviate this difficulty. Still, they are prone to spoofing and generally perform poorly in real-world scenarios (e.g., glare, occlusion, and tinted glass). This study presents TRI-GATE, a tri-modal anti-spoofing framework that unifies vehicle, license plate, and face recognition within a single, real-time decision pipeline. The system employs YOLOv4-tiny for vehicle detection and a MobileNetV2-based classifier for make–model recognition, a retrained MTCNN and LPRNet pair for license plate detection and recognition on Saudi-specific datasets (17,000 images for detection and 35,000 for recognition), and RetinaFace with InsightFace embeddings, along with a linear SVM, for driver identification. An IoU-based best-frame selection scheme reduces latency by forwarding only the most informative frame to the recognition modules. Score-level fusion is then performed by a linear SVM that learns the relative importance of each modality for the final access decision. Evaluated on a dedicated tri-modal dataset, TRI-GATE achieves 97% gate-level accuracy with an end-to-end latency of 66 ms per frame (≈ 15.15 FPS), and demonstrates robust performance in a real-world gate-like deployment, substantially improving both security and operational efficiency over existing single- and bi-modal solutions.

Author 1: Muhannad Alsultan

Author 2: Thamer Alghonaim

Author 3: Abdulaziz Alorf

Author 4: Bandar Alwazzan

Author 5: Faisal Alsakakir

Author 6: Abdullah Alhassan

Author 7: Yousif Hussain

Keywords: Tri-modal anti-spoofing; vehicle recognition; license plate recognition; face recognition; real-time gate access control; multimodal biometrics

Paper 93: Comparative Evaluation of Deep Learning Architectures and Hybrid Heuristics for Automated Gambling Content Detection

Abstract: The exponential proliferation of online gambling content represents a multifaceted challenge for contemporary automated content moderation systems, primarily driven by the sophisticated visual obfuscation and semantic complexity characteristic of modern digital advertising. This study conducts a rigorous comparative evaluation of the efficacy of Deep Learning (DL) architectures against classical Machine Learning (ML) paradigms for the deterministic identification of gambling-related imagery. Specifically, we propose and implement GADIA (Gambling Ad Detector with Intelligent Analysis), a novel hybrid funnel-based architecture that integrates structural heuristic filtering with an asymmetrically fine-tuned ResNet50 classifier. To address the systemic scarcity of high-quality public repositories, the models were trained and validated on a proprietary, strictly balanced dataset of 2,312 images, meticulously curated to encapsulate real-world adversarial marketing techniques. Performance bench-marks were established through Accuracy, Precision, Recall, F1-score, and AUC metrics. Experimental evidence demonstrates that the ResNet50 architecture attained a superior robustness profile, achieving 85.01% accuracy and 90.42% recall, significantly outperforming traditional baselines that failed to capture high-dimensional visual hierarchies. These findings validate that deep residual learning, when integrated into a hybrid heuristic-visual pipeline, provides a computationally efficient and scalable foundation for real-time platform governance and digital safety monitoring.

Author 1: Eros Anaya Sánchez

Author 2: Chesney Taichi Marchena Tejada

Author 3: Jose Alfredo Herrera Quispe

Keywords: Deep learning; image classification; gambling detection; ResNet50; hybrid systems; transfer learning; Convolutional Neural Networks; platform governance; content moderation

Paper 94: From Accuracy to Insight: Explainability in Review Rating Prediction with Transformers

Abstract: Mobile application (app) reviews provide valuable information that facilitates understanding of users’ needs, leading to better design of developed products. They have abundant data that can be utilized by different models to explain the prediction results to stakeholders. This will lead mobile app developers to trust and rely on the models that are used to develop their apps and satisfy the users’ needs. To leverage this information, outstanding improvements in complex learning algorithms have led to the development of transformer-based models that are used for natural language processing (NLP) and to exploit rating predictions. However, such models are complex and lack explainability, especially for Arabic reviews. Most studies have applied explainability models for transformer-based models to the English language and various other languages but not the Arabic language. This study presents a rating prediction explain-ability (RPE) framework that combines transformer-based and explainability models for review rating predictions from mobile government (m-government) apps. The transformer-based models predict the ratings for reviews written in English or Arabic. Then, local explainability models, such as SHapley Additive exPlanation (SHAP) and local interpretable model-agnostic explanations (LIME), explain and visualize the results. In RPE, not only high prediction accuracy was achieved for both English and Arabic reviews, but the resulted predictions were also justified with consistency between the different explainability models. The transformer-based model ELECTRA yielded the highest accuracy and F1 score of 96% for the rating prediction of English reviews, whereas the transformer-based model AraBERTv2 had 95%accuracy and F1 score for the rating prediction of Arabic reviews. The results of both explainability models provided equivalent explanations and emphasized the same words that affected the predicted ratings.

Author 1: Dhefaf T. Radain

Author 2: Dimah Alahmadi

Author 3: Arwa M. Wali

Keywords: Explainability; LIME; review rating prediction; SHAP; transformer-based models

Paper 95: Machine Learning-Based Sentiment Analysis Pipeline for Evaluating Hajj Food Service Quality

Abstract: Pilgrimage, also known as Hajj, brings together millions of people each year, creating significant challenges in managing, organizing, and maintaining the quality of various services. Among these essential services, food provision plays a vital role in shaping pilgrims’ overall experience and satisfaction. Despite its importance, research focusing on food services using sentiment analysis during Hajj remains limited. Existing studies often rely on social media data, which may not accurately capture the genuine opinions of pilgrims. This study addresses this gap by analyzing food service text reviews collected from Google Maps within the Hajj context. It contributes a new dataset collected for evaluating food services provided to pilgrims after the Hajj season, along with an empirical benchmark for Arabic Hajj food reviews. The dataset consists of 4,018 Google Maps reviews from 160 Hajj campaigns conducted between 2022 and 2025. After data preprocessing, the reviews were classified using several classical machine learning algorithms as empirical baselines, including support vector machine (SVM), logistic regression (LR), Naïve Bayes (NB), decision tree (DT), and random forest (RF). The experimental results demonstrate that LR achieved the highest accuracy of 93.6% among the evaluated models, followed by SVM and RF with accuracies of 92.9% and 92.2%, respectively. The analysis also shows that positive sentiment dominated across all studied years, indicating an overall improvement in pilgrims’ satisfaction with food services. However, the persistence of food-related issues highlights the need for continued attention and improvement in service quality.

Author 1: Amjad Enad Almutairi

Author 2: Aisha Yaquob Alsobhi

Author 3: Abdulrhman M Alshareef

Keywords: Sentiment analysis; service quality; machine learning; Hajj; food service

Paper 96: DEEP: A Distributed Energy Efficient Routing Protocol for Internet of Nano-Things

Abstract: Nanotechnology offers transformative capabilities across healthcare, environmental monitoring, and industrial automation. When integrated with modern communication technologies, Wireless Nano Sensor Networks (WNSNs) form the Internet of Nano Things (IoNT), interconnecting nanoscale devices with conventional networks. Despite its potential, efficient routing in IoNT remains challenging due to severe energy constraints, limited processing, and high propagation losses in the terahertz (THz) band. This paper proposes the Distributed Energy-Efficient Protocol (DEEP), a lightweight routing scheme designed for IoNT-based WNSNs. DEEP balances simplicity, connectivity, and sustainability through adaptive retransmission control and a hybrid energy model combining environmental energy harvesting with wireless power transfer. Performance evaluation using the Nano-Sim module of the NS-3 simulator demonstrates that DEEP significantly extends network lifetime, reduces overall energy consumption, and maintains scalability and robust delivery performance with minimal communication overhead.

Author 1: Saoucene Mahfoudh

Author 2: Areej Omar Balghusoon

Keywords: Routing protocol; Internet of Nano Things; nano-sensor; energy efficiency; energy harvesting; terahertz communication

Paper 97: Predictive Data Mining Analysis of Ownership Structures and their Influence on Corporate Tax Avoidance

Abstract: This study aims to examine the impact of corporate ownership structures on tax avoidance using a predictive data mining approach. The main challenge addressed is understanding how variations in ownership influence a firm’s strategic financial decisions, particularly its tendency to engage in tax minimization practices. By applying advanced predictive data mining techniques, the research uncovers significant patterns, identifies key ownership features, and models their relationship with tax avoidance outcomes. The dataset, derived from corporate financial statements and ownership records, is systematically preprocessed, feature-selected, and validated to ensure reliable predictive performance. Results demonstrate that differences in ownership structures significantly affect tax avoidance behavior, with certain ownership characteristics consistently emerging as strong predictors. These findings offer computational insights for both academic understanding and practical applications, helping regulators anticipate risky ownership configurations and improve policy oversight. The study highlights the importance of integrating ownership theory with predictive modeling to enhance the transparency, interpretability, and robustness of corporate tax strategy analyses.

Author 1: Tuti Herawati

Author 2: Helmi Yazid

Author 3: Nurhayati Soleha

Author 4: Munawar Muchlish

Keywords: Data mining; big data; tax avoidance; tax burden; corporate ownership structure