IJACSA Volume 17 Issue 5

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 Lifecycle-Oriented Taxonomy of Open-Source Tools for Machine Learning

Abstract: Machine learning (ML) offers various tools, frame-works and platforms for resolving complex problems in compu-tational science and engineering. Machine learning frameworks have emerged as the cornerstone of modern research and in-novation. It redefines how knowledge is produced, validated, and disseminated. Open-source machine learning frameworks are emerging as a promising way to solve the challenges of large datasets, real-time constraints and heterogeneous system components. This study provides an extensive overview of open source tools based on the ML lifecycle. These tools are evaluated based on their purpose and key features for each stage of lifecy-cle, assisting researchers and practitioners in making informed decisions according to their requirements. The key challenges are identified and future research directions are also outlined.

Author 1: Hamza Mallam Musa Mohammad

Author 2: Isa Hussain Adam Mohammed

Author 3: Abdullah Bajaber

Author 4: Syed Abdur Rahman

Author 5: Anas Sani

Author 6: Atif Naseer

Keywords: Open-source; lifecycle; taxonomy; machine learning; challenges

Paper 2: Agentic Accountability: “The Buck Stops Where?” Ethical Frameworks for Human Oversight of Autonomous AI Systems

Abstract: The rapid evolution from generative Artificial Intel-ligence (AI) toward agentic AI, systems capable of autonomously planning and executing multi-step actions, has introduced an unprecedented accountability gap in modern computing. Unlike traditional AI tools that respond to discrete prompts, agentic sys-tems pursue goals across extended time horizons, invoke external services, and produce cascading real-world consequences. This shift raises a fundamental ethical question: When an autonomous agent causes harm, who bears responsibility? This study examines the ethical and governance dimensions of agentic accountability, drawing on recent literature in AI ethics, regulatory studies, and human-computer interaction. Building on prior tiered ap-proaches to automation oversight in the human-factors literature, in regulatory risk classification, and in recent agent-autonomy frameworks, we present an action-level oversight framework that maps individual agent actions to four tiers of human-in-the-loop involvement, ranging from full automation to mandatory prohibition, calibrated by stakes and reversibility. We further analyze design patterns for “emergency brakes” (circuit breakers, action budgets, reversibility constraints, audit trails, kill switches), and propose a composition-aware extension that detects tier laundering, where individually low-tier actions compose into a higher-tier outcome. We then conduct an empirical pilot applying the framework to 177 incidents from an April 2026 snapshot of the AI Incident Database with full CSET classification, finding that 23% of tier-assignable incidents fall in T4 (prohibited under the framework) and that of 41 tangible-harm events, 26 (63%) involved Medium or High autonomy where tier-T4 enforcement would have been most directly applicable. Inter-rater reliabil-ity across the database’s three independent annotators ranges from κ=0.58 to 0.77 at the tier level (moderate-to-substantial agreement). The contribution of this work is an action-level operationalization of tiered oversight, anchored in real-world incident data, with explicit identification of composition-aware detection as the highest-leverage methodological direction for follow-up research.

Author 1: Ornella Bahidika

Keywords: Agentic AI; AI accountability; human-in-the-loop; AI ethics; autonomous systems; AI governance; responsible AI; oversight thresholds

Paper 3: ExtRA++: A Conceptual Architecture for a Deep Learning System for Aspect-Based Sentiment Analysis in User Reviews

Abstract: Aspect-Based Sentiment Analysis (ABSA) aims to identify opinion targets within textual reviews and determine the sentiment polarity associated with each target. Although transformer-based models have significantly improved contextual understanding in sentiment analysis, they remain limited in explicitly modeling structured knowledge and token-level dependencies. This study presents ExtRA++ (Enhanced Extractive Review Analysis), a conceptual deep learning architecture for fine-grained aspect-based sentiment analysis in user-generated reviews. The proposed framework integrates four complementary components: BERT-based contextual semantic modeling, adaptive external knowledge integration through Wikidata embeddings, graph-based structural reasoning using Graph Attention Networks (GATs), and sequence-consistent aspect extraction through Conditional Random Fields (CRFs) combined with aspect-aware sentiment classification. Unlike transformer-only approaches, ExtRA++ is designed as a modular systems-level architecture that combines contextual semantics, factual grounding, structural token interactions, and structured decoding within a unified framework.

Author 1: G. Kanev

Author 2: I. Valova

Keywords: Aspect-based sentiment analysis; aspect extraction; deep learning; graph attention networks; knowledge graph integration; BERT; conditional random fields; natural language processing; opinion mining; transfer learning

Paper 4: Learning Analytics for Student Performance and Early Detection of At-Risk

Abstract: The rapid growth of online learning platforms has generated large volumes of student interaction data that may support learning analytics and early academic intervention. This study proposed an intelligent learning analytics system for predicting student performance and identifying at-risk students using online learning behavior data. The Online Learning Behavior Dataset was used, consisting of demographic information, learning environment variables, and behavioral indicators. Random Forest, Support Vector Machine (SVM), Artificial Neural Network (ANN), and Gradient Boosting (XGBoost) models were implemented and evaluated using accuracy, precision, recall, F1-score, and ROC-AUC. SVM achieved the highest accuracy of 0.40, followed by Random Forest at 0.38, XGBoost at 0.35, and ANN at 0.32. However, because the task involved three risk categories, the best accuracy was only modestly above the approximate chance level of 0.33. These results indicate that the current models should be interpreted as exploratory decision-support tools rather than deployment-ready classifiers. The small performance differences among models matter for deployment because a marginal improvement may not justify automated risk classification unless supported by stronger validation, better feature engineering, and statistically meaningful performance gains. The study, therefore, demonstrates the potential of machine learning for exploratory learning analytics, while also emphasizing the need for verified institutional datasets and more rigorous evaluation before practical implementation.

Author 1: Rhowel M. Dellosa

Keywords: At-Risk student detection; learning analytics; machine learning; online learning behavior; student performance prediction

Paper 5: An EWMA-Based Adaptive Thresholding Concept for Autoencoder-Based Concept Drift Detection in Data Streams

Abstract: The static thresholds derived from the primary validation timeframe (window) are a common method for the detection of reconstruction-based concept drift. In prolonged periods of data streams, progressive changes in reconstruction accuracy frequently lead to misalignment, giving rise to repeated false alarms and inconsistency in detection behavior. This study introduces a modular lightweight adaptive thresholding strategy for the Autoencoder-Based Drift Detection Method (AEDDM) by integrating an Exponentially Weighted Moving Average (EWMA) mechanism into the batch-level decision process, without modifying the original model architecture. Rather than constituting a standalone framework, the proposed method functions as a modular enhancement to the decision layer of the AEDDM pipeline. The proposed solution is validated using a synthetic Gaussian stream together with the ELEC2 and NSL-KDD datasets. The finding demonstrates that the EWMA-based approach effectively eliminates false alarms without compromising responsiveness under abrupt changes, achieving zero-latency on NSL-KDD compared to static thresholds that produced 22 warnings and 25 false alarms in a stationary stream. Findings from this study suggest that adaptive thresholding alone significantly leads to the enhancement of detection performance in reconstruction-driven drift on a real-time stream.

Author 1: Siti Nurulain Mohd Rum

Author 2: Qiao Song

Keywords: Concept drift detection; streaming data; autoencoder; adaptive threshold; EWMA

Paper 6: Hybrid CNN for Breast Cancer Detection in Multi-Modality Imaging

Abstract: Breast cancer (BC) remains the leading cause of cancer-related death all over the world. Early accurate detection is key to the improvement of patient prognosis. The ability of advanced Artificial Intelligence (AI) methods, with a focus on Convolutional Neural Networks (CNNs), to classify breast lesions obtained from mammography and ultrasonography images is addressed in this study. Five of the latest models (ResNet-50, VGG-16, Inception-v3, custom-made CNN, and hybrid model) are evaluated using an integrated and thoroughly labeled dataset containing 10,000 images, focusing on key performance indices (KPIs), including accuracy, sensitivity, and F1-score. Furthermore, the exploration examines the challenges and protocols for integrating Explainable AI (XAI) and higher-performing models into existing clinical screening protocols and addresses issues related to trust, model generality, and ethical deployment. The findings indicate that the maximum classification accuracy (96.2%) and sensitivity of 95.8% were attained by the hybrid CNN architecture, which suggests a robust framework for safe, effective, and clinically integrated AI diagnostic support.

Author 1: Emmanuel Ofotsu Kwesi Bannor

Author 2: S. Sarah Maidin

Author 3: Vinayakumar Ravi

Author 4: Nguyen Thi Thu Thuy

Author 5: Nghiem Thi-Lich

Keywords: Breast Cancer (BC); Artificial Intelligence (AI); Convolutional Neural Networks (CNNs); Explainable AI (XAI); hybrid model (or hybrid CNN)

Paper 7: A Strategic Cloud Adoption Readiness Index (SCARI) Framework: Empirical Development and Validation in the Ghanaian Context

Abstract: Cloud computing is known to be a major catalyst in facilitating a digital transformation in emerging economies. Currently, the process of transforming into a cloud computing indigenous nation, such as Ghana, faces a series of hitherto confronting challenges, including infrastructure challenges and resistance. This study, therefore, proposes a newly developed framework known as the Strategic Cloud Adoption Readiness Index (SCARI). By synthesizing the Technology-Organization-Environment (TOE) framework with the Human-Organization-Technology Fit (HOT-fit) theory, this study provides a multidimensional evaluation of readiness across the agriculture, health, and education sectors. To validate this instrument, a mixed method that used bibliometric analysis coupled with a pilot study, “where N = 30 participants were involved”. According to the analysis, a cumulative Cronbach’s Alpha of 0.80 validated the consistency of all SCARI pillars. From the pilot findings, it is evident that despite advancements in technological complexity, human capital remains another impediment. This index provides a score that allows organizations to measure their cloud readiness while providing a forward-strategic platform to bridge the digital divide. The SCARI framework contributes a standardized benchmarking tool for assessing organizational cloud maturity and provides a strategic decision-support mechanism for managers and policymakers.

Author 1: Adade Sedom Percy

Author 2: Emmanuel Ofotsu Kwesi Bannor

Author 3: S. Sarah Maidin

Author 4: Vinayakumar Ravi

Author 5: Nguyen Thi Thu Thuy

Author 6: Nghiem Thi-Lich

Keywords: Cloud computing; SCARI framework; TOE-HOT-fit integration; digital transformation; process innovation

Paper 8: Perceived Usefulness Without Ease of Use: A TAM-Based Evaluation of Enhanced Isabela State University Echague Online Resource App for a New Guild of Education (ISUE ORANGE) Functionalities

Abstract: This study introduced a target enhancement for ISUE ORANGE, a customized learning management system (LMS) at Isabela State University (ISU)-Echague Campus, developed to address the post-2022 implementation challenges, specifically in progress tracking and exam integrity, through features such as question analytics, monitoring student reading progress, and online proctoring/examination. Modules were designed based on user-reported issues, and evaluated through the Technology Acceptance Model (TAM) with 145 students and 18 faculty using a 30-item Likert-scale questionnaire. Both groups demonstrated a strong acceptance with rating "strongly agree" on high perceived ease of use (PEOU) (grand means 4.56–4.65) and perceived usefulness (PU) (4.62–4.75), particularly for monitoring tools, indicating strong acceptance despite the atypical TAM path in this mature, localized context. Cronbach's Alpha (CA) confirmed excellent reliability (0.85–0.97), while the structural equation method (SEM) tested PU = γ₁ PEOU + ζ, revealing non-significant paths (β=0.118–0.145, p>0.05), supporting the null hypothesis (H0). The research demonstrated faculty-led innovation for ISUs or SUCs, with direct PU driving adoption over ease, offering a scalable model for hybrid learning sustainability.

Author 1: Ricardo Q. Camungao

Keywords: ISUE ORANGE; learning management system; Technology Acceptance Model; perceived ease of use; perceived usefulness; structural equation modeling; Cronbach's Alpha; question analytics; reading progress monitoring; exam proctoring; SUCs; faculty-led innovation

Paper 9: A Multilayer Secure Image Steganography Framework Using Edge-Adaptive Embedding and Pre-Encryption

Abstract: High-capacity image steganography aims to conceal large volumes of data while preserving imperceptibility and resistance to statistical and visual detection. This study proposes a multilayer secure image steganography framework using edge-adaptive embedding and pre-encryption. The method utilizes multiple edge detectors, namely Canny, Laplacian of Gaussian (LoG), and Prewitt, to accurately classify edge and non-edge regions, enabling efficient use of high-tolerance embedding areas. A Bee Colony Footprint Edge Optimization (BCFEO) algorithm is employed to select optimal embedding locations through a distortion-aware adaptive process, improving payload distribution under varying capacity conditions. For enhanced security, the secret message is encrypted prior to embedding using AES in Counter (CTR) mode, ensuring confidentiality without altering payload size and allowing exact recovery. A 5-LSB filtering mechanism is applied during preprocessing to reduce redundancy and control embedding distortion. The proposed framework is evaluated on a set of several 256×256 resized RGB images, including benchmark images from the USC-SIPI database and independently captured natural images, using standard performance metrics such as PSNR, SSIM, NCC, UIQI, and statistical steganalysis techniques. Experimental results demonstrate that the method achieves high embedding capacity with minimal visual degradation and improved performance compared to conventional edge-based approaches. The integration of adaptive optimized embedding and pre-encryption provides an efficient and reliable solution for secure image-based communication systems. The broader validation using larger datasets, different image resolutions, and more diverse image categories remains a future research direction.

Author 1: A F M Zainul Abadin

Author 2: Rossilawati Sulaiman

Keywords: Image steganography; AES-CTR encryption; hybrid edge detection; edge-adaptive embedding; BCFEO nature-inspired optimization; statistical steganalysis

Paper 10: Multimodal Machine Learning for Cybersecurity in Internet of Things Environments: A Literature Review

Abstract: This research provides a comprehensive synthesis of Multimodal Machine Learning (MML) as a transformative paradigm for IoT defense. By integrating heterogeneous data streams, including network flow statistics, device-level telemetry, and behavioral biometrics, MML architectures facilitate a holistic understanding of system states. The algorithmic advancements that were analyzed are classified into hybrid CNN-RNN structures and state-of-the-art cross-modal Transformers, and evaluate their performance across benchmark datasets such as ToN-IoT and CICIoT2023. Quantitative results show that cross-modal Transformers achieve F1-scores between 0.95 and 0.99 across detection tasks, while hybrid CNN-LSTM models range from 0.89 to 0.96. Furthermore, this study addresses the technical "optimization triad" of pruning, quantization, and edge-cloud orchestration required to deploy these models on resource-constrained hardware.

Author 1: Abdelaaziz NASSIRI

Author 2: Azeddine Khyat

Author 3: Kama El Guemmat

Author 4: Mohamed Aazi

Keywords: Internet of Things (IoT); Multimodal Machine Learning; deep learning; cyber-physical security; data fusion; Intrusion Detection Systems (IDS); Edge AI; Explainable AI (XAI)

Paper 11: Optimizing the Multi-Omics Data Types and Variant Combinations for Accurate Breast Cancer Molecular Subtypes Classification

Abstract: Breast cancer is a highly heterogeneous disease with Luminal-A, Luminal-B, HER2-Enriched, Basal-Like, and Normal-Like molecular subtypes. Accurate classification of breast cancer molecular subtypes is essential for effective diagnosis, treatment, and planning. In recent years, multi-omics data has been widely used to improve classification performance. However, most of the existing studies focus on various combinations of multi-omics data types and variants without considering their biological relevance and computational effectiveness. This research study aims to systematically analyze, validate, and optimize the combinations of multi-omics data types and variants for accurate breast cancer molecular subtypes classification. The main goal is to identify the most suitable biologically meaningful combinations for improving classification performance. This research study provides the biological rationale for integrating the multi-omics data types and variants, and analyzes the various combinations used by existing studies for breast cancer subtype classification and the reasons behind their selection. Based on this analysis, possible and best-proposed combinations of multi-omics data types and variants are presented for the accurate classification of breast cancer molecular subtypes, based on both biological and computational perspectives. In addition, this research study identifies and recommends reliable public databases that provide multi-omics datasets with verified PAM50 labels for accurate subtype classification. The findings can help researchers design more accurate and reliable classification models by using the best proposed combination of multi-omics data types and variants, and select appropriate datasets with validated subtype labels.

Author 1: Sajid Shah

Author 2: Azurah A Samah

Author 3: Syed Hamid Hussain Madni

Author 4: Sarina Sulaiman

Author 5: Zuraini Ali Shah

Author 6: Wong Yee Leng

Author 7: Aryati Bakri

Keywords: Breast cancer; molecular subtypes; multi-omics; data integration; data types and variants; datasets

Paper 12: An AI-Driven Framework for Software Effort Estimation Based on Developer Performance Metrics

Abstract: Effort estimations, including budgets, hiring people, and project timelines, in the Agile methodology, are determined by tools like COCOMO and Function-Point analysis. This study presents a framework driven by artificial intelligence (AI) that uses almost real-time signals from Git platforms that track issues, and tools to determine code quality, convert them into vectors, and trains four different regressors on them: ordinary least-squares regression, a random-forest ensemble, gradient-boosted trees, and a long short-term memory network. Hold-out evaluation together with five-fold cross-validation supplies mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination, complemented by feature-importance charts from the tree-based learners. A CI/CD-integrated retraining schedule keeps the estimator aligned with evolving team dynamics. Analyzing multi-developer projects over successive sprints reveals where predictions remain accurate and where unpredictable behavior emerges, pointing to chances for improved data gathering, enhanced governance, and more intentional feature development.

Author 1: Shaheer Ahmed

Author 2: Nosheen Qamar

Author 3: Faria Nazir

Author 4: Nosheen Sabahat

Author 5: Atif Ikram

Author 6: Najla Abdulaziz Almousa

Author 7: Hebah Abdullah Abubakr

Author 8: Mohammed Abual-Rub

Author 9: Abdulrahman Alojail

Author 10: Marwan Abu-Zanona

Keywords: Software effort estimation; machine learning; agile development; artificial intelligence; software analytics; developer performance

Paper 13: Cloud-Based Intelligent Surveillance for Digital Forensics: AI-Enhanced Criminal Investigations

Abstract: Modern smart surveillance systems have become a core element of digital forensics workflows, offering real-time detection of weapons, fire, smoke, blood, cars, individuals, and other related objects. These systems improve evidence collection, accelerate threat identification, and enhance investigative efficiency. However, most traditional surveillance architectures operate without sufficient digital forensic awareness, evidence integrity mechanisms, standardized evidence management, or reliable methods for detecting tampering. These systems rarely support chain of custody documentation and often lose reliability under adverse conditions. To address these shortcomings, a unified system that aggregates all suspicious objects and movements is needed, geared towards forensics. Therefore, we developed VigilEye, a cloud-oriented forensic surveillance framework that combines public surveillance datasets with simulated crime scene scenarios. It applies digital forensic preprocessing, such as CLAHE, to improve clarity, and SHA-256 hashing and metadata to ensure evidence integrity. Previous studies have also highlighted privacy challenges; we addressed this by sending blurred images for alerts, while the original, unblurred images are designed to be securely stored in an encrypted evidence environment accessible only to authorized personnel. This study presents the design, implementation, and digital forensic evaluation of VigilEye. The model achieved promising experimental performance (mAP50 ≈ 0.72) with an average detection speed of 33.55 ms per image, indicating that further optimization is required to enhance both speed and accuracy. The system sends alerts via the Telegram platform if it detects suspicious behavior or wanted individuals. VigilEye demonstrates the feasibility of a forensic-aware surveillance workflow with real-time alerting and integrity-preserving evidence handling in an experimental setting. Our future work includes expanding digital forensic datasets, improving spatial-temporal event detection, automating cloud-based chain of custody management, and enhancing interoperability with tools such as Autopsy and FTK Imager, alongside strengthening privacy safeguards and connected camera digital forensic tracking.

Author 1: Aseel Abdullah Aljuhani

Author 2: Fatima Hamed Aljuhani

Keywords: VigilEye; digital forensics; YOLOv8; real-time object detection; evidence integrity

Paper 14: A Fast and Efficient Residual Learning Framework Driven by Approximate Nearest Neighbor Search for Large-Scale Fingerprint-Based Visible Light Positioning

Abstract: Localization based on received signal strength using the k-Nearest Neighbor method is quite common in indoor localization systems. However, as the fingerprint dataset grows, finding the nearest neighbors becomes time-consuming. In this study, we use Approximate Nearest Neighbor (ApNN) methods to accelerate nearest-neighbor search in RSS-based localization. We further propose a residual learning framework driven by ApNN search, where ApNN provides coarse position estimates and the residual model compensates for the nonlinear relationship between RSS measurements and spatial coordinates. Simulation results show that, compared to the Brute k-Nearest Neighbor method, ApNN algorithms significantly reduce computation time. KD-Tree is the fastest algorithm, with an improvement of approximately 96% compared to kNN and WkNN. Other methods such as HNSW and Ball-Tree also achieved high performance with improvements of around 93–94%, while LSH improved by approximately 84.8%. Regarding positioning error, KD-Tree achieved the best positioning error after applying residual learning, with the highest RMSE reduction of approximately 22.5%. These results demonstrate that the proposed ApNN-based residual learning framework is an effective solution for large-scale received signal strength positioning systems.

Author 1: Huy Q. Tran

Author 2: Huu Lam Phan

Author 3: Tan Nguyen Van

Keywords: Localization; LED; approximate nearest neighbor; residual learning

Paper 15: Insight Enhancement of Distortion Effects on the Quality in Imagery Environments

Abstract: Despite advances in technology and processing, image enhancement remains an open research area. The quality of the images has a great impact on the processing, analysis, and recognition. Images are a kind of multimedia data that demand huge resources. The image quality affects further processing, and if carefully utilized, will result in a reduction in the resources required for a specific task. Despite advancements in imagery technologies, the acquisition and transmission of digital images and other factors possess some kind of distortion. This may be attributed to transmission noise, environmental effects, and poor-quality acquiring devices and cameras, to name a few. This study compares ten image enhancement techniques for three types of imagery. The images investigated represent three classes. Micro is represented by microscopic images, Macro is represented by Macroscopic satellite images, and Photography for ordinary images. The study considers compressed images, both lossy and lossless compression, and ordinary uncompressed images. Three distortion categories, mainly Noise, Blurring, and Contrast effects, are extensively experimented with and studied. Noise types investigated are Gaussian, Poisson, Salt and Pepper, and Speckle noises. Two blurring distortions, Average Blur and Gaussian Blur, are investigated. In addition to low contrast. Distortions are applied to each image, and ten enhancement filters are applied. These include Negative, De-blur Regularized, De-blur blind deconvolution, De-blur Lucy-Richardson, Contrast Stretching, Adaptive, Gray Level Slicing, Median, Digital, and Histogram Equalization. Enhancement key performance indicators, Peak Signal to Noise Ratio, and Structural Similarity Index were measured to judge the performance of the filters. The three classes of imagery experimented with a developed tool that facilitates the experimentation and gives the KPIs for each image with a friendly User Interface. Insight is gained, thus facilitating the choice for the enhancement method.

Author 1: Hanan Hassan Ali Adlan

Author 2: Mona Alanazi

Author 3: Kolood Alenezi

Author 4: Muneera Aldhabbah

Author 5: Muneera Alsubai

Author 6: Rafa Bahobail

Author 7: Reem Alharbi

Keywords: Digital image enhancement; distortion; image compression; filter; peak signal to noise ratio; mean square error; structural similarity index

Paper 16: A Hybrid Neuro-Fuzzy and Machine Learning Model for Student Performance Prediction in Online Mathematics Education

Abstract: Accurate and early prediction of student success in online mathematics education is critical for improving learning processes and developing personalized instruction strategies. However, students' problem-solving behaviors, interaction levels, and learning speeds in online learning environments inherently contain uncertainty, limiting the effectiveness of traditional assessment and singular machine learning approaches. This study proposes a hybrid neuro-fuzzy and machine learning-based student success prediction framework that combines the ability of fuzzy logic to represent uncertainty with the strong generalization and adaptive learning capabilities of machine learning methods. In the proposed approach, student success trends are primarily modeled using ANFIS, Random Forest, and XGBoost models, employing raw and derived features on the ASSISTments dataset. The predictions from these models are treated as continuous success representations reflecting uncertainty in students' learning behaviors and are used as input to a hybrid classification structure to make binary success/failure decisions. Thus, the ANFIS model is positioned as an uncertainty-aware and interpretable context generator, while the Random Forest and XGBoost models provide discriminative classification power. Experimental results demonstrate that both ANFIS and Random Forest models exhibit high individual performance; however, combining their complementary features within a hybrid structure significantly increases prediction stability and generalization. Unlike the limitations of 'black box' models in the literature, the interpretability provided by ANFIS, through its linguistic rules and membership functions, enables the proposed approach to generate pedagogically transparent and actionable implications. The findings reveal that this hybrid framework, which integrates uncertainty management with high prediction accuracy, offers a powerful decision-support mechanism for early identification of at-risk students and for developing personalized intervention strategies in online mathematics education.

Author 1: Beyza Esin Özseven

Author 2: Turgut Özseven

Keywords: ANFIS; ensemble learning; fuzzy logic; hybrid models; learning analytics

Paper 17: An Empirical Analysis of Google Play Data Safety Disclosures: A Consistency Study of Privacy Indicators in Mobile Gaming Apps

Abstract: The Google Play marketplace has introduced the Data Safety section to improve transparency regarding how mobile applications (apps) collect, share, and protect user data. This mechanism requires developers to disclose privacy and security-related practices, including data collection, data sharing, and data protection measures. However, the reliability of these disclosures depends on developer self-reporting, raising concerns about their accuracy. This study investigates the consistency between developer-reported Data Safety disclosures and observable privacy indicators extracted from Android application packages (APKs). An empirical analysis was conducted on a dataset of 41 mobile gaming apps, including 21 children-oriented and 20 general-audience apps. A static analysis approach was used to extract key privacy indicators, including device identifiers, data sharing practices, personal information access, and location access. These indicators were systematically compared with corresponding disclosures using a structured consistency evaluation framework. The results reveal varying levels of agreement across privacy categories. Device identifier disclosures show relatively high consistency (87.8%), whereas other indicators exhibit substantial mismatches. In particular, location-related disclosures show the highest inconsistency rate (56.1%), followed by personal information and data sharing indicators. Comparative analysis shows similar mismatch patterns across app categories. Chi-square tests further indicate that these differences are not statistically significant, suggesting that inconsistencies are not associated with app category but reflect broader challenges within the analyzed mobile gaming dataset. These findings highlight limitations in current marketplace transparency mechanisms and emphasize the need for improved validation approaches to ensure accurate privacy reporting.

Author 1: Bakheet Aljedaani

Keywords: Android security; data safety; mobile gaming apps; privacy analysis; static analysis

Paper 18: Improving Emotion Recognition Accuracy Using a Multimodal Model (Face and Voice Video) Based on a Convolutional Neural Network (CNN)

Abstract: Advancements in Artificial Intelligence (AI) technology have enabled the recognition of human emotions. Along with the development of deep learning and multimodal processing methods, emotion analysis can now be performed by utilizing multiple data sources simultaneously, such as facial expressions and speech signals. However, existing emotion recognition systems still face limitations in terms of accuracy. This study aims to develop and evaluate a more accurate emotion recognition system by implementing a Convolutional Neural Network (CNN)-based prediction model that integrates facial and audio data simultaneously. The study utilizes the CREMA-D dataset, which consists of visual data in the form of facial images and audio data containing variations of emotional expressions. The research process includes data preprocessing, feature extraction, and multimodal integration using an optimized Convolutional Neural Network (CNN) architecture. The evaluation results based on the F1-score indicate that the multimodal facial and audio data enable the model to recognize emotions effectively. Model performance was measured using accuracy, precision, recall, and F1-score metrics. Experimental results show that the angry (ANG) class achieved the best performance with an F1-score of 82%, while the fear (FEA) class demonstrated the lowest performance with an F1-score of only 58%. The results further indicate that the multimodal model achieved higher accuracy than unimodal models, significantly improving generalization capability on diverse testing data. This study demonstrates an overall emotion recognition accuracy improvement of 69% through the combination of facial and audio features. The analysis of combined facial and speech features on emotion classification performance shows that the proposed model achieves good overall performance, where the integration of image and audio modalities improves the correctness of facial expression predictions. Future research is expected to further improve accuracy by incorporating additional modalities beyond facial and audio data.

Author 1: Karnadi

Author 2: Ermatita

Author 3: Abdiansah

Keywords: Emotion recognition; CNN; multimodal learning; CREMA-D; face; voice; accuracy; deep learning

Paper 19: Interlayer Coupling Strategy for Two-Layer Interdependent Command-and-Control Networks: Balancing Node Walk Betweenness and Operational Attributes

Abstract: In system-of-systems operations, command-and-control networks (C2Ns) increasingly exhibit hierarchical interdependence and cross-network coupling. Designing effective interlayer coupling strategies is therefore essential for improving the survivability of interdependent C2Ns under attacks. However, most existing coupling strategies are predominantly topology-driven and fail to jointly capture multipath information transmission characteristics and operational compatibility among heterogeneous nodes. To address this issue, this study proposes an interlayer coupling strategy for two-layer interdependent C2Ns, referred to as PRO, by integrating node walk betweenness and operational attributes. Specifically, a walk-betweenness-based dependency strength metric is constructed to characterize node importance from the perspective of multipath information transmission, while an operational attribute matching mechanism is introduced to evaluate the functional compatibility of cross-layer nodes. Based on these two components, the PRO strategy is developed for survivability-oriented interlayer coupling design. Simulation experiments on an air-defense command-and-control scenario are conducted under both deliberate and random attacks. The results of repeated simulations show that the proposed strategy consistently outperforms the compared baseline methods in node survival rate, observe-orient-decide-act (OODA)-oriented functional chain retention ratio, and combat link efficiency degradation rate. Compared with the baseline random strategy, under random attack, PRO improves node survival rate by 50.00%, improves the OODA-oriented functional chain retention ratio by 19.40%, and reduces the combat link efficiency degradation rate by 3.53%. These results indicate that the PRO strategy improves not only structural robustness but also the preservation of sensing–command–firepower coordination under cascading failures. The proposed method provides a joint structural–functional framework for mitigating cascading failure risks and enhancing survivability in interdependent C2Ns.

Author 1: Bo Chen

Author 2: Yangkun Wang

Author 3: Yufeng Chen

Author 4: Yuqi Yang

Keywords: Interdependent command and control networks; interlayer coupling strategy; walk betweenness; operational attributes

Paper 20: Utilizing Structured Equation Modeling and Machine Learning in Investigating Digital Competency in Public School Teachers in Bukidnon, Philippines

Abstract: The rapid transition to digital education in the Philippines, accelerated by the COVID-19 pandemic, has highlighted significant integration challenges for public school teachers in rural provinces like Bukidnon. While digital proficiency is essential, existing studies often rely on either purely descriptive analytics or standalone machine learning models, which frequently fail to validate the complex, latent relationships between competency factors in terms of Digital Competency. To address these gaps, this research employs a two-phased hybrid analytical architecture. Phase I utilizes Structural Equation Modeling (SEM) to confirm the factor structure and establish the causal pathways of digital competency, ensuring that the theoretical framework is psychometrically sound. Phase II transitions these validated constructs into an optimized Machine Learning (ML) pipeline, incorporating SMOTE-ENN resampling to handle imbalanced regional data. Results from 1,275 participants demonstrate that "Professional Engagement" acts as the foundational engine of the digital competency system, while "Digital Pedagogy in Teaching" emerges as the most critical predictive determinant of teacher proficiency. The Random Forest algorithm achieved a high predictive accuracy of 89% and a Macro F1-Score of 85%, significantly outperforming traditional models. These findings indicate that the digital divide in this context is a pedagogical, rather than purely technical, bottleneck. The study provides a blueprint for the Department of Education to move from descriptive reporting toward Predictive Diagnostic Systems that can facilitate targeted, data-driven interventions.

Author 1: Nathalie Joy G. Casildo

Keywords: Structural Equation Modeling (SEM); Machine Learning (ML); digital competence; DigCompEdu; teacher proficiency; SMOTE-ENN; predictive analytics; educational data mining

Paper 21: DeepEdgeNet: An Edge-Cloud Deep Learning Framework for Efficient Environmental Monitoring in IoT Systems

Abstract: Deep learning (DL) is currently considered one of the most powerful tools for environmental monitoring. Many environmental variables, such as air quality, climate, water, and energy, are monitored using Internet of Things (IoT) technologies. However, the DL-based environmental monitoring systems heavily depend on the cloud, and hence they suffer from latency, high energy consumption, and data privacy. This study proposes DeepEdgeNet, a distributed DL and Machine Learning (ML) framework for environmental monitoring systems based on edge computing and federated learning. In DeepEdgeNet, the features are extracted at the edge devices, and the model updates are aggregated at the central server without sharing the sensitive data. The proposed framework was evaluated on six IoT datasets collected from various environmental monitoring systems, such as air quality monitoring, household energy consumption, satellite-based land-cover classification, climate and extreme weather analysis, water quality assessment, and drought prediction. Experimental results have shown that the proposed system significantly improves the accuracy of the considered datasets, which are 94.5% for the Air Quality dataset, 95.4% for the EuroSAT dataset, and the mean absolute error (MAE) of time-series datasets is reduced up to 0.28 for drought prediction. Moreover, the proposed system has lower inference latency, up to 130 ms, and energy consumption compared with six state-of-the-art models. Although the edge–cloud environment was simulated using a unified experimental platform, the obtained results demonstrate the effectiveness of DeepEdgeNet for scalable and privacy-preserving IoT-based environmental monitoring applications. Hence, the proposed system is efficient and applicable to IoT-based environmental monitoring systems.

Author 1: Qamar H. Naith

Keywords: Internet of Things (IoT); edge computing; federated learning; machine learning

Paper 22: Machine Learning Evaluation of the HiTar-2024 Dataset for Intrusion Detection in Smart Manufacturing Environments

Abstract: This study presents an investigation of the HiTar-2024 dataset performed in terms of the distribution of label attack types and the distribution of attacks by protocol, normal, and Denial of Service (DoS) connections over time. The investigation carried out a performance evaluation of the HiTar-2024 dataset using a machine learning approach to classify benign and malicious activities, based on BayesNet, Logistic, IBk, Multiclass, PART, and J48 classifiers. It was found that the HiTar-2024 dataset can serve as a training set for an anomaly-based intrusion detection system (IDS) in a smart manufacturing environment to detect normal and malicious activities. Furthermore, an anomaly-based IDS using the HiTar-2024 dataset is able to group malicious activities into Probing, Remote-to-Local, User-to-Root, and DoS attacks.

Author 1: Adeeb Alhomoud

Keywords: Supervised learning; intrusion detection system; security; smart manufacturing

Paper 23: ZK-FedMed: Privacy-Preserving Federated Learning for Cardiovascular and Renal Disease Prediction

Abstract: Protecting patient data confidentiality while enabling collaborative machine learning across distributed healthcare institutions remains a major challenge. This study presents ZK-FedMed, a privacy-preserving federated learning framework that combines CKKS partially homomorphic encryption for gradient protection, Rényi differential privacy with moments-accountant tracking, zk-SNARK-based integrity verification, TabTransformer feature extraction, SCAFFOLD aggregation, and an approximate federated unlearning procedure. The framework was evaluated on two public benchmark datasets: a cardiovascular disease dataset with 69,997 preprocessed records and a chronic kidney disease dataset expanded through SMOTE from 390 unique records to 4,200 balanced records. ZK-FedMed achieved 91.68% precision, 92.28% recall, and 96.73% AUC for cardiovascular prediction and 88.63% accuracy, 88.41% recall, and 94.52% AUC for renal disease classification. Because both datasets are Kaggle-derived and the CKD cohort is heavily augmented, the results are interpreted as benchmark and simulated-federation evidence rather than proof of real-world multi-hospital clinical efficacy. Additional privacy-budget sensitivity analysis showed that stricter budgets such as ε=1.0 and ε=3.0 substantially reduce utility, while ε=10.0 should be understood as a relaxed operational privacy setting rather than a strong practical privacy guarantee. The findings indicate that explicit cryptographic protection and variance-reduced aggregation can improve privacy-aware federated medical prediction while preserving clinically relevant predictive performance under clearly stated data-source and simulation limitations.

Author 1: Haewon Byeon

Keywords: Federated learning; homomorphic encryption; zero-knowledge proofs; differential privacy; transformer; cardiovascular disease; chronic kidney disease; scaffold; federated unlearning

Paper 24: An Interpretable Machine Learning Approach for Inflation Forecasting in Indonesia Using Domestic Macroeconomic Indicators

Abstract: Accurate inflation forecasting is essential for supporting forward-looking monetary policy, maintaining price stability, and preserving economic welfare. This study proposes an interpretable machine learning framework for inflation forecasting in Indonesia by integrating domestic macroeconomic indicators and seasonal variables. The study employs Random Forest, Extreme Gradient Boosting (XGBoost), and Artificial Neural Network (ANN) models using monthly data from 2008 to 2025. To systematically evaluate model performance, three experimental scenarios are implemented: baseline modeling, time-series feature augmentation, and hyperparameter optimization. Model performance is evaluated using RMSE, MAE, and R² metrics. The results consistently show that ensemble-based methods outperform ANN across all scenarios, with XGBoost achieving the best overall performance after temporal feature augmentation and hyperparameter optimization (RMSE = 0.4987, MAE = 0.3339, R² = 0.9517). The findings further indicate that temporal feature augmentation provides relatively limited improvement, whereas hyperparameter optimization substantially enhances forecasting accuracy. SHAP analysis identifies money supply (M2) and wages as the dominant contributors to inflation predictions, suggesting that monetary liquidity and labor-related factors play a more significant role than seasonal patterns in explaining inflation dynamics. Rather than contributing through algorithmic novelty, this study contributes through a systematic and interpretable forecasting framework that integrates predictive accuracy and explainable artificial intelligence to support transparent and policy-relevant inflation analysis in Indonesia and other emerging economies with similar economic structures.

Author 1: Ahmad Maimunif

Author 2: Evaristus Didik Madyatmadja

Keywords: Inflation forecasting; machine learning; explainable AI; macroeconomic indicators; time-series analysis; monetary policy

Paper 25: A Conceptual Model for Detecting Contactless Drug Distribution Based on Behavioural Analysis and Geospatial Visualisation

Abstract: This study proposes a multi-level system for detecting contactless drug distribution transactions. This system integrates behavioural pattern recognition as the primary detection channel, detection of night-time activity spikes as an enhancing module, and facial matching as an additional probabilistic evaluation layer. The system identifies a two-phase structure of covert transactions. Both the courier’s placement of the product and the buyer’s retrieval produce recognizable behavioural sequences at the same geographic location. Signal-to-noise ratio analysis identified a detection threshold at an SNR of approximately 17. It provides a quantitative foundation for camera placement planning. The behavioural pattern recognition pipeline integrates YOLO-Pose for skeleton estimation and employs classical machine learning models, including Random Forest and Gradient Boosting, for temporal action classification. The ST-GCN architecture is considered a future extension pending the availability of a larger annotated dataset.

Author 1: Medeu Kurmangali

Author 2: Talgat Akimzhanov

Author 3: Kanat Kazhibaev

Author 4: Kulambayev Bakhytzhan

Author 5: Moshkalov Altynbek

Keywords: Anomaly detection; flash detection; pose estimation; SNR; crime mapping

Paper 26: Revisiting Support Vector Machines: A Distance-Based Alternative to Deep Learning for Efficient Text Classification

Abstract: Support Vector Machines (SVMs) remain competitive in text classification, sometimes achieving comparable performance with the deep learning approach, due to their strong generalization ability and robustness to overfitting. However, their strong classification performance relies heavily on the selection of kernel functions and parameters, resulting in substantial hyperparameter tuning. A previous study has proposed Euclidean-SVM, which modifies the SVM decision mechanism by replacing the optimal separating hyperplane with a distance-based decision rule. This proposed approach reported reduced dependency on kernel functions and regularization parameters, resulting in robust performance with lower sensitivity to hyperparameter changes. Nevertheless, Euclidean-SVM only investigates Euclidean distance; other distance metrics that may achieve comparable performance remain unexplored. This study aims to evaluate the effectiveness of multiple distance metrics as an alternative decision function in the distance-based SVM framework for text classification. The distances, including Euclidean distance, Manhattan distance, Chebyshev distance, Cosine distance, and Minkowski distance, were investigated. The experimental results demonstrate that Euclidean and Cosine distances achieve stable and competitive classification performance across a wide range of hyperparameter configurations, reaching an accuracy of approximately 84-97% across the evaluated datasets. In contrast, the remaining distances, including Manhattan, Chebyshev, and Minkowski, exhibit significantly lower performance, reaching an accuracy between 14 and 71%, indicating the discriminative power of these distances is lower. A preliminary comparison with the deep learning model Long Short-Term Memory (LSTM) further shows that the distance-based SVM, including Euclidean and Cosine-based SVM, achieves higher performance and greater stability. These findings suggest that Euclidean and Cosine distances enhance the robustness of SVM-based text classification, while reducing the need for extensive hyperparameter tuning, making them suitable for resource-constrained environments compared to deep learning.

Author 1: Siew Teng Koh

Author 2: Anbuselvan Sangodiah

Author 3: Norazira Binti A Jalil

Author 4: Jafhate Edward

Author 5: Nur Fatin Liyana Binti Mohd Rosely

Keywords: SVM; text classification; Euclidean-SVM; Cosine distance; Manhattan distance; Chebyshev distance; Minkowski distance; LSTM

Paper 27: Pareto-Optimized Model Predictive Control for Dynamically Feasible Three-Dimensional Trajectory Generation in Robotic Manipulators

Abstract: This study presents a Pareto-optimized Model Predictive Control (MPC) framework for dynamically feasible three-dimensional trajectory generation in robotic manipulators operating under physical constraints. Unlike conventional interpolation-based methods that emphasize geometric smoothness while neglecting system dynamics, the proposed approach integrates a second-order discrete-time model with explicit constraints on position, velocity, and acceleration, ensuring physically consistent motion profiles. A multi-objective optimization strategy is introduced, combining grid search with Pareto front analysis to systematically tune key MPC parameters, including prediction horizon and discretization step. This enables a principled trade-off between tracking accuracy and control effort, addressing a critical limitation in existing MPC implementations that rely on heuristic parameter selection. Experimental results demonstrate that the proposed method achieves competitive tracking performance while significantly improving trajectory smoothness and reducing acceleration peaks compared to spline-based and linear interpolation approaches. The framework maintains real-time feasibility with computation times below 20 ms per control cycle, making it suitable for practical deployment in robotic systems. Furthermore, the integration of learning-based trajectory generation highlights the adaptability of the approach in complex and dynamic environments. Overall, the proposed methodology offers a scalable, interpretable, and computationally efficient solution that bridges the gap between geometric trajectory planning and physically realizable robotic motion, contributing to the advancement of control-aware trajectory generation in modern robotic applications.

Author 1: Zeinel Momynkulov

Author 2: Sayat Ibrayev

Author 3: Azizah Suliman

Author 4: Yegenberdi Tenizbayev

Author 5: Batyrkhan Omarov

Keywords: Model Predictive Control; trajectory planning; robotic manipulators; Pareto optimization; multi-objective optimization; dynamic constraints; real-time control; 3d motion planning

Paper 28: Enhanced IoT-Driven Load Forecasting with Metaheuristic-Optimized Deep Learning for Logistics Planning

Abstract: The integration of IoT technologies with smart logistics operations has opened unprecedented avenues for optimizing energy consumption in warehouse facilities. Accurate forecasting of electricity load is a key factor in cost reduction, operational efficiency, and sustainable energy management. This study presents an Enhanced Integrated Load Forecasting System (E-ILFS) that synergizes metaheuristic optimization with deep learning architectures of higher order for superior electricity load forecasting in dynamic logistics environments. Building on the foundational ILFS framework, our enhanced approach integrates Harris Hawks Optimization (HHO) for robust feature selection and an improved Residual Network (ResNet) enhanced with self-supervised learning (SSL) to more effectively capture complex, non-linear temporal patterns. Finally, comprehensive experimental evaluation on a real-world IoT-driven logistics dataset demonstrates that E-ILFS achieves state-of-the-art performance with an R² score of 0.8745, MAE of 23.59, and MAPE of 3.22%, representing a significant 12.51% improvement in R² over baseline models. In fact, the proposed system provides a practical and scalable solution for real-world logistics operations.

Author 1: Ramadan Babers

Author 2: Walid Atwa

Author 3: Mohamed Meselhy Eltoukhy

Author 4: A. M. M. Madbouly

Keywords: Electricity load forecasting; IoT; logistics planning; metaheuristic optimization; deep learning; Harris Hawks optimization; ResNet; self-supervised learning

Paper 29: A Smart IoT-Based Communication and Optimization System for Hajj and Umrah Services

Abstract: The management of Hajj and Umrah services requires reliable communication, continuous location awareness, and rapid emergency response because pilgrims move in dense and dynamic environments. This study presents a smart Internet of Things (IoT) based communication and optimization system for Hajj and Umrah services. The system integrates GPS-enabled wearable devices, a panic button emergency mechanism, cellular data transmission, a centralized cloud dashboard, and a network-aware deep neural network (DNN) component for congestion risk prediction. Unlike a purely sensing-oriented tracking system, the proposed framework explicitly incorporates communication indicators, including end-to-end latency, packet delivery ratio, jitter, and bandwidth utilization, as part of the monitoring and prediction pipeline. The prototype was implemented using a wearable tracking device and a web-based dashboard, while the prediction component was evaluated on a simulation dataset containing spatiotemporal mobility and network performance variables. Functional verification produced a black-box functional success rate of 95.6%, which is reported only as an implementation reliability indicator. The congestion risk classifier achieved 80.0% accuracy with macro-averaged precision, recall, and F1 score values of 0.72, 0.77, and 0.74, respectively. The evaluation also defines baseline comparisons, confusion matrix analysis, ROC-based assessment, and network-level stress scenarios to clarify the scientific interpretation of the results. The findings indicate that a network-aware IoT architecture can improve situational awareness and support timely operational decision-making for large-scale pilgrimage services, while full validation still requires larger real-world deployment data.

Author 1: T. H. F. Harumy

Author 2: M. Gahara

Author 3: M. F. Sipahutar

Author 4: A. Zakiyah

Author 5: N. A. Hasna

Author 6: E. Ikhsan

Author 7: Yusuf N. Buwono

Author 8: Haposan Delon Pasaribu

Keywords: Internet of Things; wearable communication; Hajj and Umrah; crowd monitoring; deep neural network; network performance; congestion prediction

Paper 30: Blockchain Consensus Mechanisms Contributing to Improved Trust in Knowledge Sharing: A Systematic Review

Abstract: Growing reliance on digital knowledge sharing across academic, corporate, and public sectors has raised serious concerns about data integrity, trust, and security. Blockchain consensus mecha-nisms offer a promising path forward through decentralized, transparent, and tamper-proof frameworks. This systematic review examines how these mechanisms enhance trust in knowledge sharing platforms, focusing on four directions: how these mechanisms are applied within knowledge sharing con-texts, the challenges they introduce for knowledge sharing de-ployment, and the advantages they provide to trust-based knowledge sharing ecosystems. Following PRISMA 2020 guide-lines, three databases Scopus, IEEE Xplore, and Web of Science were searched, and peer-reviewed studies published between 2020 and 2025 were selected for analysis. In terms of knowledge sharing applications, blockchain consensus mechanisms build trust through multiple co-occurring pathways, including distrib-uted verification, transparency, cryptographic security, immu-tability, incentive alignment, and smart contract automation. Algorithms such as Proof of Work, Proof of Stake, Delegated Proof of Stake, and Byzantine Fault Tolerance variants are widely adopted, each offering different trade-offs between secu-rity, efficiency, and scalability. In terms of challenges, scalabil-ity, energy consumption, and integration complexity with exist-ing systems remain the most significant barriers to adoption. In terms of advantages, blockchain consistently delivers stronger data security, greater transparency, and reduced dependence on centralized authorities across knowledge sharing contexts. This review concludes that blockchain consensus mechanisms offer layered and compounding trust benefits, yet technical and or-ganizational barriers continue to limit widespread deployment. Future research should focus on energy-efficient protocols, scalable architectures, and real-world effectiveness studies.

Author 1: Mohammad Fairus Bin Zulkifli

Author 2: Rabiah Abdul Kadir

Author 3: Mohamad Nazir Ahmad

Keywords: Blockchain technology; consensus mechanisms; knowledge sharing; distributed ledger technology; trust systems

Paper 31: Generative AI in Higher Education: A Systematic Review with Emphasis on Programming and Computer Science Education

Abstract: Generative Artificial Intelligence (GenAI) is rapidly reshaping teaching and learning in higher education, particularly in pro-gramming and computer science education, where tools such as ChatGPT and GitHub Copilot are increasingly used for code generation, debugging, conceptual explanation, and personalised learning support. Despite this growing use, the literature re-mains fragmented, with limited synthesis across educational outcomes, adoption contexts, user perceptions, and factors in-fluencing sustained use. To address this gap, this study conducts a systematic literature review following PRISMA 2020 guide-lines. The review draws on a dataset of 60 empirical studies published between 2022 and 2025 and retrieved from major academic databases. The selected studies were analysed using a mixed synthesis approach that combines descriptive mapping with thematic analysis. The synthesis shows that GenAI can improve coding support, debugging efficiency, conceptual un-derstanding, and student engagement, especially in program-ming-related learning contexts; however, adoption remains une-ven across regions, institutions, and course settings, while con-cerns related to academic integrity, over-reliance, reliability, and unequal access persist. By integrating findings across learn-ing outcomes, adoption patterns, user perceptions, and continu-ance-related factors, this review provides a more structured understanding of GenAI use in higher education, with particular emphasis on programming and computer science education. The study highlights the need for AI literacy, ethical governance, and inclusive institutional support to enable more responsible and sustainable GenAI integration.

Author 1: Abdulaziz Saidu Yalwa

Author 2: Mohd Shahizan Othman

Author 3: Lizawati Mi Yusuf

Author 4: Muteb Sinhat Almarshadi

Keywords: Generative artificial intelligence; higher education; programming education; computer science education; continuance intention; technology acceptance model; expectation–confirmation model; PRISMA; systematic literature review

Paper 32: Translating Job Advertisements into Competency Taxonomy: An Interpretable Approach for Robotics Recruitment Analysis

Abstract: To translate robotics recruitment texts into an interpretable competency taxonomy, this study explored how latent topics can be induced from large-scale job advertisements and used to construct a structured competency taxonomy. A domain-specific preprocessing pipeline was applied to a corpus of robotics job advertisements, and latent topics were subsequently induced using Latent Dirichlet Allocation (LDA). Building on the extracted topic evidence, a procedure was applied to organize topic summaries into task domains and competencies grounded in topic keywords. Evaluation was conducted using a set of quantitative measures to assess semantic consistency and structural quality. The results revealed recurring competency patterns encompassing on-site operation and service, engineering design and integration, software and algorithm development, and system verification and reliability assurance. The resulting competency taxonomy captures the underlying structure of employer demand and provides an interpretable basis for robotics skill demand analysis, supporting role profiling and serving as an empirical reference for Vocational Education and Training (VET).

Author 1: Zhiyan Xue

Author 2: Yang Zhou

Keywords: Competency taxonomy; robotics industry; Latent Dirichlet Allocation (LDA); Vocational Education and Training (VET)

Paper 33: Enhancing User Experience in Virtual Reality for Children with Autism: A Significant Review

Abstract: Virtual reality (VR) has become a promising technology to assist children with autism spectrum disorder (ASD especially in the areas of improving user experience with immersive, adaptive, and interactive environments. Nevertheless, current research is still disjointed in areas with very little synthesis on the effectiveness of VR in enhancing user experience and supporting the various cognitive, behavioural, and sensory requirements of autistic children. This study seeks to thoroughly review and synthesize current developments in VR-based solutions to determine the main methods, assessment plans, and design implications that can enhance user experience in this population. The systematic literature review was based on the PRISMA framework, where an advanced search strategy was implemented in Scopus and IEEE databases with keywords related to autism, user experience, and virtual reality. The identification phase yielded 939 records, followed by a rigorous screening and eligibility process resulting in a final dataset of 29 primary studies published between 2020 and 2025. The results were divided into three broad themes, namely: 1) VR-Based Intervention and Skill Development for Autism which shows how social, cognitive and daily living skills can be improved with the help of an immersive training; 2) Assessment, Evaluation and Multimodal Analytics in VR for Autism which reveals how objective evaluation and personalisation can be achieved with the assistance of behavioural and physiological and AI driven data which can improve the evaluation and personalization; and 3) Design Framework which includes VR system for Autism that focuses on users, ethical consideration and an inclusive system development. Altogether, the review indicates that VR demonstrates substantial potential to improve the user experience of children with autism, especially when it is facilitated through adaptive, data-driven, and inclusive design strategies. However, issues of scalability, long-term efficacy, and standardisation persist, meaning that future studies are required to create strong and user-friendly VR frameworks in this population.

Author 1: Nur Aleesya Mohd Asri

Author 2: Normala Rahim

Author 3: Norsuhaily Abu Bakar

Author 4: Wan Rizhan

Author 5: Ismahafezi Ismail

Author 6: Nur Saadah Mohd Shapri

Author 7: Sarah Farhana Juhari

Keywords: Virtual Reality (VR); Autism Spectrum Disorder (ASD); User Experience (UX); multimodal analytics; inclusive design

Paper 34: Credit Card Fraud Anomaly Detection in Mobile Cloud Service Security Using Extended Isolation Forest with Hyperparameter Optimization

Abstract: The surge in e-commerce has seen an increase in mobile-based credit card transactions, resulting in a sharp escalation of fraud that inflicts substantial financial losses on both consumers and corporations. Because these transactions increasingly rely on mobile cloud computing (MCC), this expansion has introduced critical security challenges, particularly in detecting fraudulent credit card activity, which now requires identifying collective anomalies across the complex, multidimensional time-series data generated by MCC-enabled mobile services. Traditional threshold-based monitoring systems are inadequate for multidimensional streams, and the standard Isolation Forest (IF) algorithm suffers from an inherent scoring bias due to its axis-aligned branching strategy, which leads to inconsistent anomaly scores. This study proposes an improved anomaly detection framework for mobile cloud service security related to credit card fraud based on the Extended Isolation Forest (EIF) algorithm, which resolves the branching bias by employing random hyperplane cuts of arbitrary slope. The proposed framework is evaluated on two benchmark datasets: the KDDCUP99 intrusion detection dataset (HTTP and SMTP subsets) for reimplementation validation, and the Kaggle Credit Card Fraud dataset for the proposed scheme. Results show that the proposed EIF achieves an AUC of 91.05%, a precision of 99.82%, a recall of 95.33%, and an F1-score of 97.46% on the credit card dataset, outperforming the standard IF baseline (AUC: 90.58%, F1: 97.35%). On the KDDCUP99 HTTP subset, the IF achieves a mean AUC of 96.21%, and on the SMTP subset, a mean AUC of 99.00% across four data shuffling runs. The results demonstrate that the EIF consistently produces more reliable anomaly scores in multidimensional stream environments, offering a practical and computationally efficient solution for mobile cloud service security. Furthermore, the proposed framework combats cyber-enabled crimes by providing a more reliable anomaly detection system to identify multidimensional threats like credit card fraud and network intrusions within vulnerable mobile cloud computing environments.

Author 1: Nur Izura Udzir

Author 2: Nur Farihin Bidin

Author 3: Aliyu Usman Shehu

Author 4: Madihah Mohd Saudi

Author 5: Azuan Ahmad

Author 6: Muhammad Harith Noor Azam

Author 7: Shazrin Azlin Ruslan

Author 8: Nor Azlinda Abdul Halim

Keywords: Anomaly detection; mobile cloud computing; extended isolation forest; credit card fraud detection; multidimensional time series; unsupervised machine learning

Paper 35: Fine-Grained Image Classification Using Vision Transformer Model

Abstract: Fine-Grained Image Classification focuses on unique features between visually similar subclasses within a wider category, which remains a challenging task due to low inter-class variations and high intra-class similarity. Conventional Convolutional Neural Network-based methods often struggle to accurately capture these minor differences. Utilizing self-attention techniques to represent global relationships within images, Vision Transformers have recently demonstrated robust performance in image classification evaluations. To enhance classification performance on complicated visual categories, this research presents a Fine-Grained Image Classification framework utilizing the Vision Transformer Model. The CIFAR-100 dataset, which includes 100 different image classes, is used for experimental purposes. The images were up-sampled because the Vision Transformer demands higher resolution inputs. To improve training efficiency and generalization, preprocessing techniques, including normalization and data augmentation, are applied. The model is trained and evaluated using standard performance metrics, including accuracy, macro precision, macro recall, and macro F1 Score, to ensure a balanced evaluation across all classes. With an overall classification accuracy of 89.68% and good macro-level assessment scores, experimental results show that the Vision Transformer Model successfully captures subtle visual distinctions among comparable categories. Transformer-based architectures offer an effective substitute for conventional techniques in Fine-Grained Image Classification applications with better performance. This research demonstrates how the Vision Transformer Model can increase classification robustness and accuracy for a dataset with very similar item classes.

Author 1: Zunaira Saleem

Author 2: Uzma Jamil

Author 3: Saman Iftikhar

Keywords: Data augmentation; fine-grained image classification; CIFAR-100; vision transformer model; deep learning

Paper 36: MRAE: Multi-Resolution Attention Ensemble with Hybrid CNN–Transformer Fusion for Breast Ultrasound Classification

Abstract: Breast ultrasound images can be classified as benign, malignant, and normal. Due to the imbalanced distribution of classes in breast ultrasound images, intra-class heterogeneity of lesions, and ultrasound artifacts like speckle noise, classification of breast ultrasound images remains a challenging problem. In this paper, we present MRAE, a hybrid architecture with a DenseNet121 convolutional encoder and two transformer encoders (ViT-Base and DeiT-Base-Distilled). These branches are run in parallel with input sizes of 192×192 pixels, 224×224 pixels, and 256×256 pixels, respectively. The learned feature representations from these branches are fused using a cross-attention block and combined using learnable ensemble weights. Focal loss with deep supervision is used during training along with CutMix regularization, Weighted Random Sampling, and Cosine Annealing. We perform experiments using 10-fold stratified cross validation on the benchmark BUSI dataset (780 Images). MRAE achieves an average accuracy, macro F1-score, and macro recall of 93.72%, 94.25%, and 95.02%, respectively, across all cross-validation folds. The ResNet50 baseline achieves accuracy, F1-score, and recall of 90.64%, 91.36%, and 91.62% across all folds. We show that MRAE has significantly lower standard deviations across cross folds, indicating better stability. Our method provides evidence that breast ultrasound images can be classified accurately and reliably in a multi-resolution attention fusion network for use in clinical breast cancer screening.

Author 1: Hemin Kareem Azeez Alshateri

Author 2: Ahmed Harbaoui

Keywords: Breast ultrasound classification; multi-resolution learning; CNN–Transformer fusion; cross-attention ensemble; Vision Transformer (ViT)

Paper 37: COAT: A Cross-Omics Attention Transformer for PAM50 Breast Cancer Subtype Classification

Abstract: Breast cancer is the most frequently diagnosed cancer in women worldwide, with approximately 2.3 million new cases annually. Accurate molecular subtyping is essential for guiding treatment decisions; however, existing PAM50 classifiers rely solely on mRNA expression and remain susceptible to normalization artifacts and platform-specific biases. To overcome these limitations, we propose COAT (Cross-Omics Attention Transformer), a novel deep learning framework that integrates mRNA, miRNA, and DNA methylation data to robustly classify PAM50 breast cancer subtypes. The model projects each omics modality into a shared latent space using modality-specific multilayer perceptrons and leverages a directed inter-omics attention mechanism to capture complementary interactions across modalities. The merged representations are processed by a classification head trained with class-weighted cross-entropy to correct for class imbalance. The model was evaluated on the TCGA-BRCA dataset (824 PAM50-tagged samples) using 5-fold stratified cross-validation, achieving an accuracy of 0.822 ± 0.020, a macro F1 score of 0.817 ± 0.033, and a macro area under the ROC curve (AUC) of 0.954 ± 0.011. These results demonstrate high performance compared to mono-omics approaches and traditional machine learning methods, while remaining competitive with recent multi-omics models. An additional 10-fold cross-validation experiment with Bayesian hyperparameter optimization further improved performance (accuracy = 0.852 ± 0.030, macro F1 score = 0.836 ± 0.041, macro AUC-ROC = 0.957 ± 0.013), indicating stable performance across different validation conditions. GradientSHAP interpretability analysis revealed that COAT identified biologically relevant biomarkers, including ERBB2 and GRB7 for the HER2-enriched subtype, ESR1 and PGR for the Luminal A subtype, and KRT5 and FOXM1 for the Basal subtype. Overall, COAT demonstrates that directed inter-omics cross-attention effectively integrates complementary multi-omics signals, achieving strong predictive performance while preserving biological interpretability and providing a generalizable framework for multi-omics-based cancer classification.

Author 1: Soufiane El Atfa

Author 2: Abdelmajid Hajami

Author 3: Hamid Machhour

Author 4: Hakim Allali

Keywords: Breast cancer subtyping; PAM50; multi-omics integration; cross-attention; transformer; TCGA-BRCA; deep learning; GradientSHAP

Paper 38: Efficient Vulnerability Classification in IoT Networks: An Approach Using Convolutional Neural Networks and Tabu Search Optimization

Abstract: In this study, researchers propose a novel solution for efficient enhancement of vulnerability detection in several IoT environments. Efficient Vulnerability Classification has been introduced as the presented technique in IoT Networks (EVCIN). The method, EVCIN, which is a proposed approach, utilizes CNNs with Tabu Search Optimization. Customized CNN models have proven to be extremely accurate in identifying vulnerabilities for IoT classes, showing half (6 layers), Sefunten 99.03% (7 layers), and 95.71% (8 Layers). The contribution of using Tabu Search did increase the accuracy of classification through introducing an effective set of techniques that head towards the optimal solutions. Throughout the study, the superior performance of EVCIN was demonstrated in characterizing vulnerabilities when it was compared against single CNN and Tabu Search models and state-of-the-¬art methods. Data visualization and AUC analyses were also effective for understanding the performance and discrimination ability of models. There are numerous important implications from the study of EVCIN for enhancing cybersecurity in IoT and also adding vitality to the development of vulnerability classification in IoT networks. The above approach gives a potentially useful solution in a reliable and efficient way for vulnerability finding. This would then enhance security and flexibility in IoT-based networks.

Author 1: Feras Fares AL-Mashagba

Author 2: Mohammad Othman Nassar

Author 3: Essam Said Hanandeh

Keywords: Vulnerability categorization; IoT; networks; convolutional neural networks; tabu search optimization; cyber security I

Paper 39: Scoping Review on Global Digital Policies and Transformation Strategies for Ageing Societies: Implications for Malaysia

Abstract: The rapid growth of ageing populations presents major social, healthcare, and economic challenges worldwide. Governments are increasingly adopting digital transformation strategies to support ageing societies through artificial intelligence (AI), the Internet of Things (IoT), telehealth, smart environments, and interoperable data systems. However, existing studies often focus on individual technologies, single-country initiatives, or sector-specific programmes, resulting in a fragmented understanding of how digital policies address ageing populations at the global level. This study conducts a scoping review to map digital policies, frameworks, and transformation strategies related to ageing societies and to identify transferable implications for Malaysia. The review follows PRISMA-ScR guidance and applies a Population-Concept-Context (PCC) framework. Searches were conducted across Scopus, Web of Science, IEEE Xplore, ACM Digital Library, PubMed, and institutional repositories for publications and policy documents from 2015 to 2025. After screening 857 records, 63 sources of evidence were included. The findings show that global digital ageing policies cluster around five domains: healthcare digitalisation, social inclusion and active ageing, smart environments and age-friendly cities, governance and data regulation, and economic participation of older adults. Persistent challenges include digital literacy gaps, privacy and ethical concerns, fragmented governance structures, limited interoperability, and unequal digital infrastructure. This review provides a consolidated global overview and highlights policy mechanisms that can inform the development of integrated digital ageing frameworks, particularly for countries preparing for population ageing, such as Malaysia.

Author 1: Ahmad Wiraputra Selamat

Author 2: Shamsul Arrieya Ariffin

Keywords: Ageing nation; digital framework; ageing society; digital transformation; scoping review

Paper 40: A Mid-Level Feature Fusion Framework Integrating PHLF and VGG16 for Robust Batik Pattern Detection

Abstract: Javanese batik is characterized by distinctive motif patterns. The rapid evolution of designs through motif combination has introduced increased complexity, posing challenges for the community. People are increasingly less familiar with traditional Javanese original batik patterns. Many studies using CNN have been conducted to recognize batik patterns. However, it is important to improve object detection performance by strengthening explicit local features. To answer this question, our study aims to improve the detection of batik patterns that have local texture patterns and complex orientations that are in harmony with batik geometry with an integrated fusion scheme. To improve detection, an integrated fusion schema model was developed using the PHLF hybrid framework as a geometrically oriented local feature with VGG16 as a deep feature extractor for object detection. According to research, although VGG16 is reliable on large benchmarks such as ImageNet, VGG16 is less reliable for subtle intra-motif variations, which can suppress accuracy. Evidence from the results of the study shows that in the eight-class dataset consisting of 6,400 images for training data and 1.600 images for test data, a mid-level feature fusion approach on VGG16 with integrated PHLF improves resistance to data variations such as lighting, fabric deformation, and background complexity. The experimental results showed that the model achieved an mAP value of 0.68 at IoU = 0.5 and 0.2846 at IoU = 0.7. The significant difference between mAP@0.5 and mAP@0.7 suggests that the model still has limitations in the precision of the localization of the boundary box.

Author 1: Jani Kusanti

Author 2: Edi Noersasongko

Author 3: Purwanto

Author 4: Moch Arief Soeleman

Keywords: A Mid-level feature; detection; fusion framework; phlf; traditional javanese batik

Paper 41: YOLO-CBAM: A Lightweight Attention-Guided Deep Learning Framework for Real-Time Road Damage Detection

Abstract: Accurate and real-time assessment of road infrastructure is critical for smart city maintenance and transportation safety. However, conventional object detection models often struggle with complex environmental factors, such as varying illumination, shadows, and background noise, leading to false detections and missed fine-grained defects. In this study, we propose YOLO-CBAM, a lightweight and fast neural network architecture tailored for real-time road surface damage detection. The standard YOLO11s backbone is enhanced through the integration of a Convolutional Block Attention Module (CBAM), which synergistically applies channel and spatial attention mechanisms. This modification enables the network to actively suppress irrelevant background visual noise and focus exclusively on structural defects like longitudinal cracks and potholes. Extensive experiments conducted on a comprehensive dataset reveal that the implementation of partial transfer learning significantly mitigates early-stage gradient shock, allowing the model to achieve a mean Average Precision (mAP@50) of 0.60 in just 40 training epochs. Deployed on an NVIDIA RTX 4070 Ti, the proposed framework achieves an inference speed of 25 frames per second (FPS), demonstrating an optimal balance between detection accuracy and computational efficiency. The YOLO-CBAM model provides a robust, cost-effective solution for automated video surveillance and road condition monitoring in smart city infrastructures.

Author 1: Olzhas Olzhayev

Author 2: Bakhytzhan Kulambayev

Author 3: Azizah Suliman

Author 4: Assel Rustem

Author 5: Almira Madiyarova

Author 6: Batyrkhan Omarov

Keywords: Computer vision; deep learning; object detection; YOLO architecture; CBAM; attention mechanism; road monitoring; transfer learning

Paper 42: Deep Learning-Based Recommender System for Arabic Content with Integrated Sentiment Analysis of User Reviews

Abstract: Recommender systems are widely used as an information filtering technology to automatically predict and identify a set of interesting items for users based on their needs and preferences. They are widely applied in many domains, including e-commerce, social media, education, and healthcare. Recommender systems employ various filtering approaches, such as collaborative filtering, content-based filtering, and hybrid methods. Collaborative filtering is broadly categorized into memory-based and model-based approaches. Deep learning-based recommenders are a type of model-based approach that employs neural networks to capture patterns in user preferences and item features and generate accurate and personalized recommendations. In this study, we apply deep learning-based recommender systems to the Large-Scale Arabic Book Reviews Dataset (LABR) and evaluate their performance. To improve recommendation quality, we integrate sentiment analysis of user reviews using pre-trained Arabic BERT–mini and AraBERT, enabling more accurate modeling of user preferences. The results show that the combination of deep learning techniques and sentiment analysis produces more accurate recommendations, improving user satisfaction and engagement with Arabic content.

Author 1: Amani Al-Ajlan

Author 2: Nada Alshareef

Keywords: Recommender system; deep learning; sentiment analysis; LABR dataset; Model-based collaborative filtering; pre-trained model Arabic BERT; pre-trained model AraBERT

Paper 43: A Knowledge-Enhanced Cross-Modal Transformer Network for Sentiment Analysis in Intelligent Interaction

Abstract: With the rapid advancement of multimodal emotion recognition technology, sentiment analysis models that integrate heterogeneous information—such as facial expressions and vocal intonation—are driving human–computer interaction and affective computing toward multidimensional, objective, and highly accurate approaches. Conventional emotion recognition methods typically rely on a single-modal input and therefore struggle to capture complex semantic associations and deep emotional features, which in turn undermines the stability of recognition results. A knowledge-enhanced cross-modal Transformer network (KCTN) model was proposed for sentiment analysis, which incorporates a multimodal fusion module and a long-range affective integration module to achieve deep collaborative modeling across text, speech, and facial expression features. This framework substantially enhances the completeness and robustness of emotional semantic representations. Experimental results on the self-built EC-SFED multimodal dataset and the publicly available dataset CMU-MOSI demonstrate that KCTN surpasses several mainstream baseline models in both accuracy and macro-averaged F1 score, validating its superior performance in intelligent interaction and affective computing applications.

Author 1: Chunyan Huang

Author 2: Xinlu Sun

Keywords: Multimodal sentiment analysis; transformer network; emotion recognition; psychological assessment; intelligent interaction

Paper 44: Enhancing Traffic Congestion Forecasting with Explainable Deep Learning: A Framework Using LIME for Transparent Intelligent Transportation Systems

Abstract: Intelligent transportation systems aim to improve traffic management and road safety, manage traffic effectively, and reduce roadway system congestion. This optimally requires estimating future traffic congestion. Unfortunately, the most popular machine learning and deep learning techniques can be unsuitable for model development in this task due to interpretability challenges. This study attempts to provide a solution to this challenge by creating a tool that integrates Local Interpretable Model-agnostic Explanations (LIME) into any traffic congestion forecasting system. This tool is applied to the Metro Interstate Traffic Volume dataset, which contains samples of traffic and road system congestion along with temporal, weather, and contextual data. For global feature analysis, a Random Forest Regressor is used as a baseline model, while a neural network model is developed to predict the congestion of the traffic and road system. The neural network model achieved a congestion prediction with an R² score of 0.612, a mean squared error of 0.026, and a mean absolute error of 0.129. The LIME tool also provides temporal feature insights, which show that examples of weekday/holiday status reduce the sample congestion prediction for the example, while precipitation increases it. At a global level, hour of the day, day of the week, temperature, and month of the year are the dominant factors in congestion prediction. These findings illustrate the value of adding interpretability to predictive models of traffic congestion when using explainable artificial intelligence.

Author 1: Ouhmidou Hajar

Author 2: Nabou Abdellah

Author 3: Elikram Moulay Ahmed

Keywords: Explainable artificial intelligence; lime; traffic congestion forecasting; intelligent transportation systems; random forest

Paper 45: Mammographic Image Classification Based on KAN-CBAM

Abstract: With the fast enhancement of deep learning, research on automatic detection of breast tumors is becoming increasingly in-depth. However, traditional CNNs’ linear kernel has difficulty not only in capturing the nonlinear combination of low-frequency structures and high-frequency details but also in fully exploring the nonlinear discriminative features in medical images. Furthermore, the current models used for breast tumor detection have complex structures and slow inference speeds. Therefore, this study solves the linear kernel problem and improves the model inference speed by using a lightweight mammographic image classification method based on KAN-CBAM to speed up breast cancer diagnosis. The proposed method introduces the KAN convolution module, which embeds a learnable B-spline activation function into the convolution kernel. This scheme improves the capability of the proposed method to capture nonlinear features and improves its capacity to fit complex, nonlinear distributions. Moreover, the proposed method combines the CBAM attention mechanism to screen key semantic channels through channel attention and then uses spatial attention to locate lesion areas, achieving "channel- space" dual feature recalibration, further improving the attention to key features, and achieving more accurate classification in complex and variable medical images. We evaluated the proposed method on the mammographic image datasets DDSM, INbreast, and MIAS to verify its performance. The results prove that KAN-CBAM models have higher adaptation to diverse dataset scales, efficiently acquiring major lesion parts and nonlinear discriminatory features in mammographic images. Meaningful and great enhancements were seen in different metrics such as accuracy, F1-score, AUC, precision, and recall, demonstrating extensively improved model strength and generalization capability.

Author 1: Yuanyuan Wang

Author 2: Vladimir Mariano

Keywords: KAN; CBAM; deep learning; mammographic; classification

Paper 46: Enhancing Virtual Team Performance Through a Quantitative-Based Management Model (VTPM)

Abstract: The rapid growth of globalization, digital transformation, and remote working practices has significantly increased the reliance on virtual teams within modern organizations. Despite their advantages, virtual teams continue to face challenges such as communication barriers, inadequate technology management, unclear role allocation, ineffective strategic planning, and excessive documentation practices, all of which negatively affect team performance. This study aims to identify the key factors affecting virtual team performance and to develop a structured Virtual Team Performance Management (VTPM) model to improve the effectiveness of virtual collaboration. A quantitative research approach was adopted, involving questionnaire-based data collection with 110 respondents who had worked in virtual teams (instrument reliability testing was conducted using Cronbach’s Alpha analysis). The findings indicate that role clarity, effective communication, effective strategic planning, task allocation, and appropriate management of technology are important in the determination of the performance of virtual teams. Specifically, 39.1% of respondents rated communication as “needs improvement,” 63.6% agreed that excessive documentation reduces efficiency, and 56.4% strongly agreed that strategic planning is effective. On this premise, the Virtual Team Performance Management Model (VTPM) has been constituted and validated by three academic and industry experts through expert evaluation. The model received positive evaluations from academic and industry experts regarding its practicality and relevance in virtual team environments. This study provides a comprehensive and empirically informed model that organizations can apply to improve the work of virtual teams.

Author 1: Sathya A/L Nantha Kumar

Author 2: Zulkefli Mansor

Keywords: Virtual teams; team performance; remote work; VTPM model; communication; quantitative study

Paper 47: An Explainable XGBoost-Based Framework for Robust Multi-Cohort Prediction of Pancreatic Cancer

Abstract: Pancreatic cancer remains a leading cause of cancer-related mortality due to its asymptomatic progression and late-stage diagnosis. Early detection is critical for improving patient prognosis and clinical outcomes. Traditional diagnostic approaches and previous computational models often struggle with molecular heterogeneity and technical variations across different genomic platforms. These batch effects limit the reliability and generalizability of predictive biomarkers when applied to diverse clinical settings. This research proposes a robust machine learning framework designed for platform-invariant pancreatic cancer prediction. Large-scale transcriptomic datasets, including microarray data from the Gene Expression Omnibus (GEO) and RNA-seq data from The Cancer Genome Atlas (TCGA), were integrated. Subsequently, the ComBat algorithm was applied to correct batch effects. This resulted in a discovery cohort of 441 samples and an external validation set of 409 samples. An optimized XGBoost classifier was developed through comparative benchmarking. It was compared against several learners, including Random Forest, LightGBM, Support Vector Machines (SVM), and Logistic Regression. The model demonstrated high predictive performance, achieving an internal test AUC of 0.923. External validation was performed across six independent cohorts, yielding a mean AUC of 0.761 ± 0.090 (95% CI: 0.689–0.833). These findings support the robustness and cross-platform generalizability of the proposed framework. To enhance model interpretability, SHapley Additive exPlanations (SHAP) analysis was employed to identify key molecular drivers. These drivers were further validated using biological enrichment analysis through Over-Representation Analysis (ORA) and log2FC-weighted Gene Set Enrichment Analysis (GSEA). The proposed framework provides a reliable and scalable solution for multi-platform integration. This approach facilitates accurate risk stratification and precision oncology in clinical practice.

Author 1: Nada Ahmed El-Gammal

Author 2: Rania Ahmed Abdel Azeem Abul Seoud

Author 3: Sayed T. Muhammad

Keywords: Pancreatic cancer; gene expression analysis; XGBoost; SHAP explainability; pathway enrichment; Explainable AI (XAI)

Paper 48: A Hybrid PCA–Random Forest Model for Predicting Employee Performance from Work Ethics and Work Values

Abstract: This study examined the predictive relationship of work ethics and work values with employee task performance using a hybrid PCA–Random Forest model. Data were obtained from 231 PSU employee respondents and 81 questionnaire variables covering task performance, work ethics, and work values. A quantitative, cross-sectional predictive design was used; therefore, the findings are interpreted as predictive associations rather than causal effects. Descriptive statistics, reliability analysis, correlation analysis, Principal Component Analysis (PCA), baseline regression models, Random Forest regression, and a hybrid PCA–Random Forest model were applied. PCA reduced the 74 predictor items into 19 components, explaining 80.9% of the predictor variance. On the held-out test set, the hybrid PCA–Random Forest model achieved MAE = 0.272, RMSE = 0.383, and R² = 0.524. Standard Random Forest produced a similar test performance (MAE = 0.283, RMSE = 0.383, R² = 0.526), indicating that PCA did not substantially improve accuracy but produced a more compact and less redundant feature representation. Professionalism, commitment to public interest, and nationalism and patriotism emerged as important predictors of task performance. The study demonstrates the usefulness of interpretable machine-learning models for evidence-based human resource development in public higher education, while noting limitations related to ceiling effects, self-report data, single-institution sampling, and the need for external validation.

Author 1: Nova E. Arquillano

Keywords: Employee performance; work ethics; work values; principal component analysis; random forest; hybrid model; machine learning; PSU employees

Paper 49: A New Approach for 3D Shape Retrieval Based on an Explainable Boosting Classifier

Abstract: In the past decade, the number of available 3D models has grown rapidly. This growth is mainly driven by the development of scanning devices and increasing demand for 3D meshes in many application fields. As a consequence, it becomes crucial to have consistent content-based retrieval systems of large 3D mesh repositories. However, many existing methods still struggle to achieve a good compromise between efficiency and accuracy, especially on large and heterogeneous databases. In this work, we propose a supervised framework for building compact but discriminative 3D shape descriptors. For each mesh, we start by computing three features - the dihedral angles between adjacent faces, the Shape Diameter Function (SDF), and the Shape Index - then we convert them into normalized histograms and concatenate them into a single feature vector, which is then used as input to an Explainable Boosting Classifier (EBC). After training, the classifier produces, for each mesh, a short probability vector that we use as its numerical descriptor. Experiments on the standard Princeton Benchmark database validate our approach, achieving a mean Average Precision (mAP) of 97.23%, outperforming the selected baseline methods under the adopted experimental protocol.

Author 1: Fatima Rafii Zakani

Author 2: Mohcine Bouksim

Author 3: Khadija Arhid

Author 4: Taoufiq Gadi

Author 5: Mohamed Aboulfatah

Keywords: 3D shape retrieval; content-based indexing; shape descriptor; explainable boosting classifier

Paper 50: HBC-Glicko: A Bloom-Constrained Adaptive Assessment Architecture with Uncertainty-Aware Progression

Abstract: Adaptive assessment systems typically model learner ability as a single continuous latent trait and select items accordingly. Although effective for efficient measurement, such models rarely encode hierarchical cognitive progression as a formal constraint within the learner model itself. In educational domains where higher-order performance depends on sufficiently stable prerequisite knowledge, this omission may permit pedagogically incoherent trajectories, including premature advancement or misleading readiness judgments produced by compensatory aggregation across cognitively distinct levels, where strength at one level can mask weakness at another. This study pursues three objectives: 1) to formalize an adaptive assessment architecture in which Bloom’s revised taxonomy operates as a non-compensatory structural constraint on learner progression, meaning that readiness must be established at each prerequisite level independently, without any area of strength substituting for insufficient evidence elsewhere; 2) to derive the principal theoretical properties that follow from this design; and 3) to situate the proposed architecture relative to existing rating-based learner models. To address these objectives, the study proposes Hierarchical Bloom-Constrained Glicko (HBC-Glicko), a theory-driven measurement architecture for formative adaptive assessment. Instead of representing learner state as a single scalar estimate, HBC-Glicko models it as a band-specific vector with band-level uncertainty. Progression is regulated through anchor-based readiness thresholds and confidence-bound decision rules, such that advancement depends on both estimated performance and evidential stability at prerequisite levels. The study formalizes the learner model, within-band routing logic, threshold construction, promotion and prerequisite reinforcement rules, and derives the principal architectural properties. The contribution is conceptual and architectural rather than empirical, establishing a formally specified foundation for subsequent simulation and empirical investigation.

Author 1: Wissal EL Fougour

Author 2: Mohamed Erradi

Keywords: Adaptive assessment; Bloom's revised taxonomy; Glicko rating system; learner modeling; non-compensatory progression; uncertainty modeling; formative assessment

Paper 51: A Secure Chaotic DNA-Based Framework for Satellite Image Encryption

Abstract: With the growing use of satellite imaging in environmental monitoring, defense, and remote sensing applications, protecting satellite images during storage and transmission has become increasingly important. Unlike textual data, satellite images contain high spatial redundancy and strong correlations among neighboring pixels, which can limit the effectiveness of conventional encryption methods when applied directly. To address this issue, this study presents a chaotic DNA-based framework for satellite image encryption. The proposed approach combines logistic-map-based chaotic key generation with DNA-inspired encoding and XOR operations to enhance pixel-level confusion and diffusion. The method was evaluated on a grayscale satellite image using common statistical security measures, including information entropy, adjacent-pixel correlation, NPCR, and UACI. The reported results indicate that the encrypted image achieved high entropy, low correlation between adjacent pixels, and strong sensitivity to small changes in the input image. These findings suggest that the proposed framework provides a reasonable basis for secure satellite image encryption. Although the current evaluation is limited in scope, the method shows encouraging performance and offers a useful direction for further investigation on larger and more diverse satellite image datasets.

Author 1: Asim Seedahmed Ali Osman

Author 2: Ibrahim Rizqallah Alzahrani

Author 3: Aeshah Khalil I Alotaibi

Author 4: Adil Mahmoud Mohamed Mahmoud

Author 5: Daifallah Zaid Alotaibe

Keywords: Chaotic cryptography; DNA computing; satellite image security; image encryption; information security; secure image transmission

Paper 52: A Diffusion-Based Generative AI Framework: An Exterior House Design from Textual Descriptions

Abstract: The architectural design process is often iterative, time-consuming, and heavily dependent on effective communication between clients and professionals. Existing design tools, such as Computer-Aided Design (CAD) systems, require technical expertise, limiting accessibility for non-professional users. This study proposes a generative artificial intelligence framework for exterior house design using a diffusion-based text-to-image model. The proposed approach integrates Stable Diffusion for image generation with a vision-language model (BLIP) to enhance semantic alignment between textual descriptions and generated outputs. In addition, an interactive refinement mechanism based on image inpainting is incorporated to allow localized modification of design elements. The system is trained on a dataset of exterior house images and evaluated using quantitative metrics, including CLIP Score and Fréchet Inception Distance (FID), as well as usability assessment. Experimental results demonstrate that the proposed framework is capable of generating semantically relevant and visually coherent architectural designs, while improving accessibility and reducing the time required for design iteration. The findings highlight the potential of generative AI as an effective tool for supporting user-centric architectural visualization and design exploration.

Author 1: Muhammad Amirul Akmal bin Ajusin

Author 2: Noor Hasimah Ibrahim Teo

Author 3: Rosniza Roslan

Author 4: Raseeda Hamzah

Author 5: Anita Ahmad Kasim

Keywords: Generative artificial intelligence; stable diffusion; text-to-image generation; architectural design; image inpainting

Paper 53: Integration of Random Forest and Hough Transform for Cancer Classification Using Microarray Gene Expression Data

Abstract: Cancer classification poses a significant challenge owing to the intricate nature and diversity of the disease. This study introduces a novel methodology for cancer classification leveraging microarray gene expression data. The proposed approach integrates Random Forest (RF) and Hough Transform (HT), where RF performs feature selection and classification, and HT identifies sub-biclusters that are merged into larger clusters using a hypergraph model. Evaluation on multiple cancer datasets demonstrates that the hybrid approach improves classification accuracy compared to standalone RF or HT while capturing meaningful gene expression patterns.

Author 1: Hibah Alatawi

Author 2: Hechmi Shili

Keywords: Cancer classification; microarray gene expression; random forest; Hough transform; feature selection; hypergraph model

Paper 54: Detection of Unauthorized Use in SIEM Through Behavioral Analysis and Adaptive Rules

Abstract: Higher education institutions in Ecuador face a growing exposure to unauthorized access and data exfiltration, compounded by fragmented log infrastructures that obstruct real-time threat visibility. This study addresses those gaps through the design and deployment of a Security Information and Event Management (SIEM) architecture at the Compu Sur Higher Technological Institute (ITECSUR), augmented with User and Entity Behavior Analytics (UEBA) and a context-sensitive adaptive rule engine. Rather than relying on static signature matching, the proposed model constructs individual behavioral baselines per user and asset, dynamically escalating alert thresholds according to geographic context, access time, and asset sensitivity classification. Empirical validation conducted over 450 security events, including simulated Salgorea Trojan injections, supply chain compromise scenarios, and government-grade spyware indicators, yielded a Mean Time to Detect (MTTD) reduction from 48.5 hours to 12.4 minutes (99.57%), a recall rate of 95%, and a 65% decrease in false positives relative to rule-only baselines. Hardening protocols applied in parallel reduced exposed network ports by 78% and elevated institutional compliance with Ecuador's Organic Law on Personal Data Protection (LOPDP) from 35% to 92%. The architecture, built on AlienVault OTX and Osquery agents, processed over 1.2 million daily Indicators of Compromise autonomously, demonstrating operational feasibility for institutions with constrained IT budgets. These findings position SIEM-UEBA integration as both a technical countermeasure and a regulatory compliance instrument for the higher education sector.

Author 1: Julio Armando Landázuri Castro

Author 2: Renato M. Toasa

Author 3: Maryory Urdaneta Herrera

Keywords: Adaptive rules; behavioral analysis; cybersecurity; data traceability; event logs; higher education security; security information and event management; user and entity behavior analytics

Paper 55: Semantic Modeling of Medical Specialty Relationships Using Large Language Models

Abstract: This work proposes a computational framework for modeling semantic relationships between medical specialties using large language models. Forty-four medical specialties officially recognized in France were analyzed using Claude 4 Sonnet, GPT-4.1, and LLaMA 3.2 3B. Each model evaluated the relevance of 307 ICD-11 disease families, 260 educational teaching items, and 276 technical skills. From these ratings, criterion-specific similarity matrices were constructed and aggregated into composite matrices. The framework includes hierarchical clustering, substitution-coverage analysis, Mantel correlation tests, adjusted Rand index evaluation, and heatmap-based visualization of inter-model differences. Claude 4 Sonnet and GPT-4.1 produced highly consistent similarity structures, with a mean off-diagonal similarity of 0.867, a standard deviation of 0.045, and strong matrix correlation. LLaMA 3.2 3B generated more homogeneous patterns, indicating reduced differentiation while preserving global structure. Hierarchical clustering revealed five stable groups of specialties aligned with functional medical domains. At similarity thresholds above 0.90, most specialties had two to five semantically close candidates, suggesting a basis for exploratory analysis of short-term cross-specialty coverage under appropriate expert and institutional constraints. These results suggest that large language models can produce stable and interpretable representations of relationships between medical specialties. The proposed framework provides a data-driven approach for analyzing specialty proximity and can support exploratory applications in medical education structuring, cross-specialty coordination, and health-system planning.

Author 1: Ismail Bouajaja

Author 2: Omar Elfahim

Author 3: Omar Bouattane

Author 4: Oussama Barakat

Author 5: Abdelaziz Daaif

Keywords: Large language models; semantic similarity; computational modeling; cluster analysis; medical specialties

Paper 56: An Intelligent Semantic-Aware Academic Library System Using Natural Language Processing and Knowledge Graphs

Abstract: Traditional library systems designed for academic research suffer from poor integration when employing AI-based techniques for enhancement purposes. These limitations are mainly related to ignoring semantic relationships during search, overlooking hidden relationships among words, and lacking the generalization capability required to enhance researchers’ experience. To address these issues, the Intelligent Library System (IntLS) is proposed and further enhanced through effective knowledge graph modeling to enable more accurate retrieval of results. In addition, the basic NLP processing steps are optimized to capture hidden relationships during tokenization, stop-word removal, and stemming stages. The architecture of the proposed system consists of two main components: a semantic component, responsible for generating semantic representations of documents, and an analytics component, responsible for analyzing historical searches to predict future user needs and support effective resource management. The proposed system is compared with two well-known systems using similarity-based accuracy and a set of AI-based evaluation metrics. The enhanced system, Enh-IntLS, demonstrates superior performance, achieving a 1.9% improvement in similarity-based accuracy and a 1.7% improvement in AI-based accuracy.

Author 1: Mohamad Shady Alrahhal

Author 2: Bandar Abdulwahed Fallatah

Author 3: Omar F. Aloufi

Author 4: Abdullah M. Barakeh

Keywords: Library systems; academic search; generalization; knowledge graph; similarity; accuracy

Paper 57: Rice Pest Detection Using Enhanced YOLOv8n with Multilayer Contextual Attention and Deformable Snake Convolution

Abstract: Accurate and intelligent detection of rice pests is critical for ensuring food security and advancing precision agriculture. However, due to the small size, irregular morphology, dense distribution, and complex backgrounds of pest targets, traditional lightweight detection models often suffer from low recall and poor localization accuracy in real-world paddy field environments. To address these challenges, this study proposes an enhanced detection model, YOLO-Rice, based on the YOLOv8n framework. First, a Multilayer Contextual Attention (MLCA) module is embedded after the SPPF layer to collaboratively fuse channel, spatial, local, and global contextual information, thereby enhancing the model's sensitivity to subtle pest features. Second, the original C2f structure is redesigned into C2f-DS, which integrates Dynamic Snake Convolution (DSConv) to improve adaptive perception of deformable pest contours and irregular morphological edges. Finally, the conventional CIoU loss is replaced with a WIoU loss to guide the network to focus more effectively on hard-to-fit small and occluded targets. Extensive experiments on a self-constructed dataset of 11 common rice pest species demonstrate that YOLO-Rice achieves 84.8% Precision, 69.9% Recall, 78.7% mAP@0.5, and 63.4% mAP@0.5:0.95, representing significant improvements over the baseline YOLOv8n model. The proposed approach achieves an excellent balance between detection accuracy and computational efficiency, making it highly suitable for real-time deployment on UAVs and edge devices in agricultural pest monitoring applications.

Author 1: Shuangyuan Li

Author 2: Jianglong Lin

Author 3: Jiaming Liang

Author 4: Tianyu Li

Keywords: Rice pest detection; YOLOv8n; deep learning; real-time detection

Paper 58: Data-Driven Excimer LiDAR Framework for Joint Surface Reflectivity Mapping and Atmospheric Pollutant Profiling

Abstract: This study proposes a data-driven excimer LiDAR framework for joint surface reflectivity mapping and atmospheric pollutant profiling, integrating physics-based sensing with deep learning-based multi-source data fusion. The system utilizes ultraviolet excimer LiDAR measurements in combination with auxiliary data from UAV platforms, satellite observations, and ground-based sensors to construct a unified environmental monitoring pipeline. A structured signal processing approach is applied to extract physically meaningful features, including backscatter and extinction coefficients, as well as differential absorption parameters. These features are subsequently fused using a deep learning architecture designed to model complex nonlinear relationships across heterogeneous data sources. Experimental results demonstrate that the proposed method achieves high predictive accuracy, with improved correlation and reduced error compared to traditional LiDAR and baseline fusion approaches. The framework effectively captures both vertical atmospheric pollutant distributions and horizontal surface reflectivity patterns, enabling comprehensive environmental analysis. Validation against external datasets confirms the robustness and generalization capability of the model under varying conditions. The integration of data-driven modeling with excimer LiDAR sensing enhances system performance while maintaining real-time capability. Overall, the proposed approach provides a scalable and efficient solution for advanced environmental monitoring, contributing to the development of intelligent remote sensing systems for air quality assessment and land-cover analysis.

Author 1: Sandugash Dospanbetova

Author 2: Gulzat Ziyatbekova

Author 3: Murat Baktybayev

Author 4: Botakoz Smagul

Author 5: Yermakhan Zhabayev

Author 6: Zhanar Bidakhmet

Keywords: LiDAR; data-driven modeling; multi-source data fusion; deep learning; remote sensing

Paper 59: Cervical Cytology Classification Using Multiple CNN Architectures with Transformer-Based Feature Enhancement

Abstract: Cervical cancer is the fourth most common cancer among women worldwide and remains a major public health challenge, particularly in regions with limited access to early screening and diagnosis. Accurate classification of cervical cytology images is critical for early detection of cervical cancer, which remains a major health burden in low- and middle-income countries. This study presents a comprehensive evaluation of multiple deep learning architectures for the automated classification of Pap smear images into four diagnostic categories: Negative for Intraepithelial Lesion or Malignancy (NILM), Low-Grade Squamous Intraepithelial Lesion (LSIL), High-Grade Squamous Intraepithelial Lesion (HSIL), and Squamous Cell Carcinoma (SCC). We systematically compare eight Convolutional Neural Network (CNN) architectures: AlexNet, VGG-16, GoogLe-Net, Network-in-Network (NIN), ResNet-50, DenseNet-121, Capsule Networks, and EfficientNet-B0 on a publicly available cervical cytology dataset. To enhance feature representation and capture long-range dependencies, we additionally incorporate a Vision Transformer (ViT-16) model. All models are trained and evaluated under identical preprocessing and sampling conditions to ensure fair benchmarking. Experimental results demonstrate that ViT-16 achieves the highest test accuracy of 95.88% and an overall specificity of 0.9864, outperforming all CNN counterparts. EfficientNet-B0 and DenseNet-121 also showed strong performance, achieving 94.33% and 93.30% accuracy, respectively. Notably, ViT-16 provided superior classification outcomes for challenging minority classes such as SCC and HSIL. The findings highlight the growing potential of transformer-based models in cytopathology and underscore the importance of architectural design in developing robust diagnostic tools. This work contributes a comparative foundation for future research in AI-assisted cervical cancer screening systems.

Author 1: Mehreen Sirshar

Author 2: Omama Shakeel

Author 3: Nayyab Asim

Author 4: Fakeeha Jafari

Author 5: Hani Almoamari

Author 6: Adnan Nadeem

Author 7: Mohammad Zubair Khan

Author 8: Ibrahim Aljubayri

Keywords: Cervical cancer; CNN architectures; vision transformer

Paper 60: Methodology for Identifying Soiling on PV Panels Using RGB Images and Deep Learning

Abstract: Soiling is one of the major factors that can affect the performance of PV installations in many regions of the world. On the other hand, the variety of PV models, types of soiling, and climate conditions makes it very difficult to create a universal model. To address this problem, this study presents a methodology for the identification of soiling on PV panels via semantic segmentation, which can support the decision-making process in terms of surface cleaning and automation of the process. The methodology includes data collection, preparation of training and testing data, training of models, and application of the optimal one. Next, the methodology is demonstrated using a small dataset of clean and dirty PV panels, three neural network architectures (DeepLab v3, U-Net, and PSPNet), and two backbone models (ResNet34 and ResNet50). The obtained results show the feasibility of the methodology and allow highlighting the DeepLab v3 model with a ResNet34 backbone as the best-performing algorithm for identifying pigeon droppings. The second-best combination is the U-Net + ResNet34, which showed good efficiency for identifying smaller dirty areas. The proposed methodology could be useful for operators of large-scale photovoltaic installations by supporting the decision-making process when it comes to the timely cleaning of specific areas for performance improvement and lower costs.

Author 1: Katerina Gabrovska-Evstatieva

Author 2: Tsvetelina Kaneva

Author 3: Irena Valova

Author 4: Dimitar Trifonov

Author 5: Nikolay Valov

Author 6: Ventsislav Keseev

Author 7: Nicolay Mihailov

Author 8: Boris Evstatiev

Keywords: PV soiling; semantic segmentation; pixel-based classification; neural network; RGB images

Paper 61: Gas-Efficient Smart Contract Design: Quantifying Refactoring Impact on EVM Execution Costs

Abstract: High transaction costs remain a major barrier to the scalability of Ethereum-based decentralized applications (DApps), particularly when smart contracts are computationally inefficient. Although the Solidity compiler optimizer can reduce bytecode size and improve some low-level patterns, it does not fully address structural inefficiencies in storage layout and state mutation. This study introduces controlled empirical research on the topic of manual smart contract refactoring approaches with the aim of quantifying their impact on gas usage and execution cost in the Ethereum Virtual Machine (EVM). The Remix Integrated Development Environment (IDE) and a synchronized Go-Ethereum (Geth) node (version 1.13.5) were configured to create a controlled experimental environment. This environment was connected to the Sepolia Testnet to approximate conditions similar to the Ethereum Mainnet. The role of high-cost storage operations such as SSTORE was analyzed using opcode-level transaction traces, which were collected using debug_traceTransaction. The proposed refactoring plan implies the alignment of storage slots by systematically packing the variables and data location optimization (calldata and memory) to minimize unnecessary memory allocation. The experiments show gas reductions of up to 40.68% for storage-intensive functions, with an average reduction of 28.5% across all evaluated test cases. Moreover, the findings at the opcode level have shown that it is possible to reduce the costs of unnecessary storage writes without impacting the correct functional performance of the execution. Overall, the findings show that storage-aware manual refactoring is a viable strategy for improving runtime efficiency and reducing the execution cost of Layer-1 smart contracts.

Author 1: Nur Haliza Abdul Wahab

Author 2: Juniardi Nur Fadila

Author 3: Nur Faszha Razali

Author 4: Keng Yinn Wong

Keywords: Ethereum smart contract; solidity refactoring; gas optimization; Ethereum Virtual Machine (EVM); opcode tracing

Paper 62: Agent-Oriented Fuzzy Decision Support System for Multi-Criteria Evaluation of Sustainable Investments in the Agro-Industrial Sector

Abstract: Sustainable development of the agri-food sector in emerging economies requires the use of analytical tools capable of taking into account climate risks, environmental constraints, and investment flow instability when making management decisions. Given the fragmentary nature of statistical information and the high volatility of the external environment, traditional econometric methods for assessing investment attractiveness demonstrate limited effectiveness and low interpretability. This study proposes an agent-oriented modular fuzzy decision support framework for the comprehensive assessment of sustainable investments in the agricultural sector. The developed approach combines a modular data processing architecture that provides automated collection and preprocessing of heterogeneous statistical sources (OECD, FAO, and national statistics), with a fuzzy additive aggregation (Fuzzy-SAW) mechanism that allows for interpretable multi-criteria assessment of economic, environmental, and production-forecasting factors. The methodological novelty of the study lies in the integration of an automated data processing pipeline with an explainable fuzzy multi-criteria assessment model focused on conditions of data incompleteness and structural uncertainty. Empirical validation of the model was performed using statistical data from the agro-industrial complex of the Republic of Kazakhstan for the period 2010–2023. The results show that the proposed framework effectively smooths out short-term volatility in indicators and identifies long-term structural trends in investment attractiveness. In particular, in 2021–2023, the integral index of sustainable investment remained at around 0.37, despite adverse climate shocks, mainly due to the compensatory effect of growth in private investment flows, which indicates the formation of mechanisms for the adaptive sustainability of the agricultural sector. The proposed analytical framework is a scalable and interpretable decision support tool that can be used by government agencies, investors, and industry analysts in developing long-term strategies for sustainable agricultural development in emerging economies.

Author 1: Aigerim Omurtayeva

Author 2: Ulzhan Makhazhanova

Author 3: Akmaral Kulamanova

Author 4: Dinara Kargabaeva

Author 5: Bolat Tassuov

Author 6: Adilbek Tanirbergenov

Keywords: Sustainable investments; agro-industrial sector; decision support system; agent-oriented systems; fuzzy logic; ESG factors; climate risks; emerging economies; multicriteria evaluation

Paper 63: Phonetic Completeness Over Prosodic Diversity: Syllable-Level Synthetic Corpus Construction for Low-Resource Penang Hokkien Speech Synthesis

Abstract: This study presents the first Text-to-Speech (TTS) model for Penang Hokkien, a low-resource tonal dialect at risk of extinction. To address phonological sparsity in the collected speech corpus, we propose a two-stage fine-tuning approach that emphasizes comprehensive phonetic coverage through syllable-level synthetic augmentation while subsequently refining prosodic naturalness using real speech recordings. By supplementing a limited 45-minute real speech corpus with a 2-hour syllable-level concatenative synthetic corpus, the full dialectal inventory of approximately 2,000 unique syllable-tone combinations was encompassed. Experimental results suggest that improving syllable-tone coverage contributes substantially to intelligibility and tonal accuracy in this low-resource tonal setting. Technical optimizations, including a 600-ms cross-fading technique to mitigate boundary artifacts and numerical tone markers to reduce token sparsity, further improved model stability and synthesis quality. The final model achieved a Mean Opinion Score (MOS) of 3.92.

Author 1: Yu Liang Lai

Author 2: Yen Min Jasmina Khaw

Author 3: Seng Poh Lim

Author 4: Tien Ping Tan

Keywords: Low-resource language; speech synthesis; text-to-speech; data augmentation techniques

Paper 64: Exploring New Possibilities: Virtual Reality in Supporting Children with Autism

Abstract: Children with autism spectrum disorder (ASD) often require structured opportunities to practice social interaction, following instructions, and recognizing emotions in safe, controlled environments. This study presents and evaluates a mobile virtual reality (VR) application designed to train these skills through simulated everyday scenarios. An applied quantitative study with an experimental design involving control and experimental groups was conducted with 10 children with ASD from the Sagrada Familia Special Basic Education Center (Trujillo, Peru). The experimental group (n=5) used a Unity 3D/C# VR application featuring four scenarios (park, store, pedestrian crossing, and emotion recognition), while the control group (n=5) followed conventional activities. Outcomes were collected through direct observation (60 records; 30 per group) and analyzed statistically, with findings interpreted cautiously due to the repeated-observation structure of the data. The post-test results showed statistically significant differences between groups, with the experimental group performing better on all evaluated indicators (social interaction, following instructions, emotion recognition, and average time to detect oncoming traffic). However, due to the small sample size and repeated observations, these findings should be interpreted as preliminary evidence. The results suggest that mobile VR may complement traditional interventions to strengthen social and emotional skills in children with ASD.

Author 1: Brayan J. Lozano-Martel

Author 2: Eduardo F. Araujo-Vasquez

Author 3: Segundo E. Cieza-Mostacero

Keywords: Autism Spectrum Disorder; virtual reality; mobile application; social skills training; immersive learning; special education

Paper 65: Augmented Reality in Automotive Technical Training: Technological Innovation Through Mobile Applications in Trujillo

Abstract: This study evaluated the impact of an Augmented Reality (AR)-based mobile application on technical training in automotive maintenance at a company in Trujillo, Peru. The application was developed using Unity 3D, the Vuforia marker-based recognition engine, and C#, enabling technicians to interact with three-dimensional models of automotive components — including engines, valve trains, crankshafts, and braking systems — through touch-based rotation, zoom, and perspective controls overlaid on physical QR markers. This marker-based interaction architecture distinguishes the system by enabling step-by-step guided visualization of complex maintenance procedures without requiring physical equipment access. A posttest-only control group experimental design was employed with 60 technicians randomly assigned to experimental and control groups, using the Mobile-D methodology for software development. The results revealed a significant improvement in technical content comprehension, a 46.6% reduction in average training time (from 10,862 to 5,784 seconds), and increased satisfaction levels, reaching a mean score of 4.0 versus 2.6 in the control group. The Mann–Whitney U test confirmed statistically significant differences (p = 0.001) across all indicators, with large effect sizes (Rosenthal's r ≥ 0.86). These findings suggest that AR-based mobile applications represent a viable instructional approach for automotive maintenance training, though broader generalization requires validation across diverse organizational contexts.

Author 1: Julissa B. Polo-Holguin

Author 2: Eduardo F. Araujo-Vasquez

Author 3: Segundo E. Cieza-Mostacero

Keywords: Augmented reality; marker-based AR; automotive maintenance training; mobile application; technical comprehension; training efficiency; experimental design

Paper 66: CFD-Based Aerothermal Design Verification of Coffee Cabinet Dryer

Abstract: Traditional sun drying remains the primary practice among smallholder coffee farmers in the Philippines, often leading to an inefficient and inconsistent drying process. While transitioning to mechanical drying improves product quality, achieving uniform airflow in cabinets with multiple trays often requires costly post-fabrication modifications, resource-intensive trials, and repetitive field testing to fine tune the performance of the system. To reduce inefficient and repetitive trial-and-error procedures, a pre-fabrication design verification of a small-scale mechanical cabinet dryer was conducted using COMSOL Multiphysics. This study explores how the dryer’s geometry influences its aerodynamic and thermal performance under baseline, no-load conditions prior to fabrication. The system was simulated to deliver an inlet air velocity of 1.56 m/s and a controlled heat flux to maintain 40oC. Results revealed steady pressure equalization within the plenum chamber, which successfully mitigated non-uniform airflow. As a result, a relatively high Velocity Uniformity Index of 0.84 was achieved across the tray layers. In addition, internal turbulent mixing contributed to a more stable thermal profile, reaching a near-ideal Temperature Uniformity Index of 0.99. These findings reveal that the core geometry of the dryer performs well from an aerodynamic standpoint. By achieving flow equalization in a no-load setting, the study provides an optimized baseline design, helping reduce the need for extensive trial-and-error during future prototyping and testing under actual operating conditions.

Author 1: Gerry M. Castillo

Author 2: Aeron R. Mojica

Author 3: Jhon Heron P. Carsocho

Author 4: Ray Rojan C. Romeroso

Author 5: Sheryl Dinglasan-Fenol

Author 6: Al Eugene L. Torres

Author 7: Gee Jay C. Bartolome

Keywords: COMSOL Multiphysics; coffee dryer; heat transfer; design verification; aerothermal simulation; performance characterization

Paper 67: Unveiling Temporal Dynamics of Consumer Emotions in Online Reviews: An Emotion Mining Approach Using Ekman’s Emotion Model

Abstract: Currently, business organizations are using Electronic Word-of-Mouth (EWOM) from consumer-opinion platforms to elevate their marketing strategies using semantic analysis and emotion mining. Existing efforts are devoted to analyzing EWOM using fine-grained emotion classification; however, the well-known and well-established emotion models that provide a basic set of psychological and biological emotions are not targeted in these studies. Further, the existing literature is unable to present temporal dynamism and volatility of consumers’ emotions according to emotion categories that have an inescapable impact on consumers and corporate decision-making process. Therefore, this study tries to address this gap by extracting, classifying, summarizing, and tracking consumers’ emotions for corporate and consumers’ decision-making processes based on widely used and well-established Ekman's emotions model. Therefore, the current work aims to provide consumers’ emotions temporal dynamism and volatility at product and feature levels by using binary and fine-grained emotion classifications based on Ekman’s basic psychological and biological emotions. Online reviews from amazon.com are used for experimental purposes. The results of the study exhibited that consumers predominantly expressed joy emotion to the camera. However, the picture feature of the camera evoked negative emotions, namely, fear, anger, and sadness. Temporal analysis unveiled shifting emotional expressions in quarter 1 and quarter 2. The results of the study will assist corporates in shaping their decision-making process and marketing strategies. Moreover, the findings of the study provide valuable insights into consumers to improve their purchase decisions by pinpointing the advantages and disadvantages of critical product features through emotional clues.

Author 1: Azra Shamim

Author 2: Abdulwahab Ali Almazroi

Keywords: Emotion; emotion mining; semantic analysis; online reviews

Paper 68: Privacy Preserving Federated Graph Learning with Data Envelopment Analysis Driven Interpretable Customer Segmentation Framework

Abstract: Customer segmentation plays a critical role in retail analytics by enabling personalized marketing, optimized resource allocation, and data-driven strategic decision-making. However, customer data is often distributed across multiple retail branches and contains sensitive transactional information, creating significant challenges related to privacy preservation, regulatory compliance, and model interpretability. Traditional segmentation approaches, including clustering algorithms and centralized deep learning methods, typically require aggregated data storage, which increases privacy risks and limits secure deployment in distributed retail environments. In addition, many existing methods fail to capture complex relational patterns such as co-purchasing behavior and inter-customer structural dependencies. To address these limitations, this study proposes FedGraph-DEA, a novel hybrid framework integrating federated learning, graph neural networks (GNNs), and Data Envelopment Analysis (DEA) for privacy-preserving and efficiency-aware customer segmentation. The framework first employs Distributed Federated Convolutional Autoencoders to extract latent customer representations from decentralized retail datasets. Similarity graphs are then constructed locally, followed by federated GNN-based community detection to identify structurally coherent customer groups without sharing raw data. Finally, DEA is applied to evaluate the operational efficiency of the discovered customer segments. Experimental evaluation was conducted using the UCI Online Retail dataset partitioned across five simulated non-IID client nodes. The proposed model achieved 96.1% accuracy, 0.95 precision, 0.96 recall, and 0.95 F1-score when compared with pseudo-ground-truth clusters generated through K-Means reference clustering. Furthermore, the framework obtained a silhouette score of 0.74 and a modularity value of 0.57, demonstrating strong cluster compactness and structural separation. The proposed system provides a scalable, interpretable, and privacy-preserving solution for distributed retail analytics.

Author 1: Pentareddy Ashalatha

Author 2: G. Krishna Mohan

Keywords: Customer segmentation; federated learning; graph neural networks; data envelopment analysis; privacy preservation

Paper 69: Explainable Cognitive Graph Intelligence Framework for Multi-Task Graduate Employability and Salary Prediction

Abstract: Accurate prediction of graduate employability and expected salary is essential for enabling timely, data-driven career interventions in academic institutions. However, most existing methods rely on standalone machine learning or deep learning models that fail to jointly capture cognitive traits, temporal academic progression, and peer-level relational structures within a unified and interpretable framework. This study introduces ExCogGNet, a hybrid multi-task learning architecture that integrates Multi-Layer Perceptron (MLP), Bidirectional Long Short-Term Memory (BiLSTM), Transformer, and Graph Neural Network (GNN) components, optimized using Bolas Spider Optimization (BSO). The model simultaneously predicts student placement outcomes and salary levels while incorporating cognitive trait awareness. Static psychometric attributes are encoded via MLP, sequential academic trajectories are modeled using BiLSTM, contextual feature interactions are captured through Transformer-based self-attention, and relational dependencies among students are learned using Graph Attention Networks (GAT). The BSO algorithm optimizes hyperparameters, attention weights, and fusion coefficients under a unified multi-task objective. A key contribution is a counterfactual SHAP-based explainability module that converts feature attributions into actionable, personalized recommendations for improving employability and skill readiness, enabling prescriptive educational decision support. Experimental results on the Campus Recruitment dataset show 96% accuracy and 0.97 AUC for placement prediction, along with an RMSE of 24,000 INR and R² of 0.91 for salary estimation. The model outperforms baseline methods including SVM, Random Forest, XGBoost, CNN, BiLSTM, Transformer, and standalone GNNs, with statistical significance confirmed via McNemar’s test (p = 0.003), demonstrating strong predictive and interpretability performance.

Author 1: Hameeda Khatoon

Author 2: Chandra Prakash Vudatha

Keywords: Cognitive properties; prediction of placement; salary regression; graph neural network; bolas spider optimization; explainable AI; MLP-BiLSTM-Transformer

Paper 70: Large Language Models for Arabic Automated Essay Scoring

Abstract: Automated Essay Scoring (AES) has become an important research area in educational artificial intelligence due to its potential to support scalable and consistent assessment. The developments within the realm of transformers and large language models (LLMs) have led to great improvements within AESs through their ability to comprehend semantics and context, as well as their knowledge of rubrics. Although these advancements have been realized within the English language, there is still relatively little research surrounding Arabic Automated Essay Scoring (AAES). This survey summarizes some of the latest advancements in AAES and discusses traditional ML models, deep learning, transformers, and LLM-driven evaluation frameworks. In this study, researchers synthesize the relevant literature regarding datasets, prompts, pre-processing methods, performance metrics, and reliability of scores. Typical performance metrics used to analyze the level of agreement between human raters and automated systems are QWK, MAE, and correlation-based metrics. The survey also describes crucial challenges encountered by AAES systems such as insufficient amount of data, inconsistencies, high computation costs, bias, and non-reproducibility. Overall, it can be said that both transformers and LLMs achieve better performance when it comes to capturing context information and providing agreement with human assessment. However, issues with reproducibility and scalability continue to persist. Additionally, the survey presents new areas of research that may be relevant to future AAES studies, such as multilingual evaluation, hybrid grading, explainability, and standardized sources of Arabic essays.

Author 1: Leena Najjar

Author 2: Liyakathunisa Syed

Keywords: Arabic automated essay scoring; large language models; educational AI; prompt engineering; evaluation metrics; natural language processing

Paper 71: Revolutionizing Urban Waste Systems: A Comprehensive Review of IoT, AI, and Optimization Techniques for Smart Waste Management

Abstract: Urban waste management is shifting from fixed, reactive collection toward data-driven and adaptive service models. This review synthesizes 33 recent studies and technical contributions on Artificial Intelligence of Things (AIoT) for smart waste management. The synthesis is organized around six analytical dimensions: IoT sensing and communication, real-time monitoring performance, dynamic routing, AI-based classification and robotic sorting, edge/fog/cloud intelligence, and circular governance. Unlike a descriptive survey, the paper develops an AIoT-SWM taxonomy and an evaluation rubric for comparing smart waste systems according to interoperability, latency, energy profile, scalability, decision autonomy, circular-economy contribution, and social inclusion. Reported deployments show measurable operational gains, including 26-35% reductions in fuel consumption, more than 30% improvement in fleet utilization, 25% gains in collection efficiency, and up to 48% reduction in overflow incidents. AI sorting studies also report controlled classification accuracies above 90%, while recent techno-economic evidence indicates 95.1% material purity, 50 items/minute throughput, and payback periods of 4.3-4.9 years under specific emerging-economy conditions. The review concludes that AIoT can improve municipal waste services only when technical performance guarantees are combined with open data standards, cybersecurity safeguards, human oversight, and context-sensitive inclusion of informal waste actors.

Author 1: Lahcen GOUSKIR

Author 2: Abdelmoula ABOUHILAL

Author 3: Mohamed GOUSKIR

Author 4: Mohamed BASLAM

Author 5: Soufiane BELHOUIDEG

Author 6: Hanaa Hachimi

Keywords: Smart waste management; Internet of Things; Artificial Intelligence; AIoT; edge computing; dynamic routing; circular economy; robotic sorting

Paper 72: Deep Learning-Enhanced Excimer Laser LiDAR System for High-Resolution Earth Surface Monitoring and Multi-Platform Validation

Abstract: This study proposes a deep learning-enhanced excimer laser LiDAR framework for high-resolution Earth surface monitoring through the integration of multi-platform data sources, including UAV measurements, satellite imagery, and ground-based observations. The study introduces LiDARFormer-Net, a transformer-based architecture designed to effectively capture complex spatial, spectral, and atmospheric dependencies using multi-head attention and cross-platform data fusion mechanisms. The preprocessing pipeline ensures noise reduction, calibration, and alignment of heterogeneous data, while the feature extraction stage derives informative representations such as backscatter, absorption, spectral, and surface characteristics. Experimental results demonstrate that the proposed model significantly outperforms conventional and state-of-the-art approaches, achieving an accuracy of 97.32%, an RMSE of 0.053, an MAE of 0.042, and a coefficient of determination of 0.983. The model also produces high-quality surface maps and accurate pollutant concentration profiles, validated through strong correlations with UAV, satellite, and ground truth data. Ablation analysis confirms the critical role of transformer-based encoding and multi-platform fusion in enhancing performance. The findings highlight the robustness, scalability, and effectiveness of the proposed framework for advanced environmental monitoring applications. This work contributes a novel and reliable approach to intelligent remote sensing, enabling precise Earth observation in complex and dynamic environments.

Author 1: Sandugash Dospanbetova

Author 2: Gulzat Ziyatbekova

Author 3: Murat Baktybayev

Author 4: Botakoz Smagul

Author 5: Yermakhan Zhabayev

Author 6: Zhanar Bidakhmet

Keywords: Deep learning; LiDAR; transformer architecture; earth surface monitoring; environmental monitoring; remote sensing; feature extraction; attention mechanism

Paper 73: Federated and Cross-Domain Student Performance Prediction

Abstract: Accurate student performance prediction is critical for data-driven educational decision-making; however, it is often hindered by data heterogeneity, privacy constraints, and domain shift across academic contexts. This study investigates student final grade (G3) prediction using three complementary machine learning paradigms: centralized learning, cross-domain generalization, and federated learning with personalization. Experiments were conducted on Portuguese and Mathematics student datasets using traditional regression models, ensemble methods, neural networks, and a personalized federated learning framework based on FedProx and FedBN. In the centralized setting, models capable of capturing non-linear relationships, particularly XGBoost and multi-layer perceptrons, achieved superior predictive performance, with XGBoost attaining an R2 of 0.8308 and the lowest error metrics. In contrast, direct cross-domain application of models trained on Portuguese data to Mathematics outcomes resulted in severe performance degradation, with several models yielding negative R2 values, highlighting the adverse impact of domain shift. To address privacy and heterogeneity challenges, a federated learning simulation was implemented. While the global federated model achieved moderate accuracy, the introduction of local personalization led to substantial performance gains. The personalized client models achieved stronger local predictive performance than the global federated model and showed competitive performance relative to centralized baselines. Learning-curve analysis further indicate that model performance in centralized settings improves with increasing data size but eventually plateaus, whereas cross-domain learning remains constrained despite additional data. In federated learning, predictive performance consistently im-proves across training rounds, demonstrating the effectiveness of iterative collaboration and client-level personalization. Overall, the results suggest that federated learning with personalization offers a competitive privacy-preserving alternative to centralized modeling and provides a clear improvement over direct cross-domain transfer in heterogeneous educational analytics.

Author 1: Sam Zhe Xuan

Author 2: P. Ganesh Kumar

Author 3: C. Rani

Author 4: Kanagalakshmi

Author 5: R. RajiniGanth

Author 6: Atif Mahmood

Keywords: Federated learning; cross-domain generalization; educational data mining; student performance; quality education

Paper 74: Intelligent Traffic Surveillance: Machine Learning-Based Detection of Vehicle Speed Violations

Abstract: With the rapid increase in the number of vehicles on roads, traffic management, and safety enforcement have become significant challenges worldwide. Traditional speed violation detection systems either employ high-end hardware, expensive computational resources, or post-processed video data, which are inefficient to implement in real time. This study presents a real-time intelligent vehicle speed violation detection system using YOLOv8 for object detection and SORT for vehicle tracking, and a new Speed Detection Algorithm (SDA). The system can effectively detect vehicles and calculate their speed from video recorded by low-cost fixed cameras. Unlike other models that process simulated or post-processed video data, the new model processes real-life scenarios such as changing lighting and weather conditions. Experimental results indicate that the system achieves 92% to 95% vehicle detection accuracy while maintaining a Mean Absolute Error (MAE) of 1.8 km/h and Root Mean Square Error (RMSE) of 2.5 km/h for speed estimation, and 98% effective at speed detection compared to various other systems that came before it in terms of real-time processing effectiveness. This cost-effective and scalable solution can be incorporated into traffic observation systems for the improvement of road safety and regulation of speed limit compliance.

Author 1: Niloy Kanti Paul

Author 2: Dipanwita Saha

Author 3: Kaushik Biswas

Author 4: Sultanul Arifeen Hamim

Author 5: Tanvir Ahmed

Author 6: Rifath Mahmud

Keywords: Machine learning; speed detection; object detection; vehicle tracking; violation

Paper 75: Malak: A Python Toolkit for Edge AI Model Optimization

Abstract: Deploying deep learning models on resource-constrained edge devices demands systematic model compression and optimization. PyTorch supplies low-level quantization and pruning primitives, yet a typical quantization-aware training workflow still requires approximately 40 to 60 lines of observer wiring, calibration, and conversion boilerplate, which slows iteration for researchers and students. We present Malak, an open-source Python toolkit that wraps these primitives behind a small task-oriented application programming interface exposed through both a Python module and an edgeai command-line interface. The toolkit covers the full prototyping loop: model training; post-training quantization (dynamic and static) and quantization-aware training; magnitude and structured pruning; knowledge distillation; export to the Open Neural Network Exchange format; per-layer latency profiling; and a Kullback–Leibler-divergence drift check for deployed models. We empirically validate the com-pression subset (post-training quantization, quantization-aware training, pruning, and knowledge distillation) on the CIFAR-10 and Fashion-MNIST image-classification benchmarks with five architectures: MobileNetV2, ResNet18, ResNet50, EfficientNet-B0, and a custom convolutional network we refer to as Sim-pleCNN. Across three random seeds, quantization-aware training yields a 3.48× reduction in model size on MobileNetV2 with accuracy statistically indistinguishable from the 32-bit floating-point baseline (77.91±0.83 per cent versus 77.68±0.92 per cent). Dynamic post-training quantization preserves accuracy within 0.13 per cent across all tested architectures, and magnitude pruning at 50 per cent sparsity holds within roughly one percentage point of the baseline after three fine-tuning epochs. A knowledge-distillation experiment confirms that the toolkit reproduces the qualitative behavior of Hinton-style soft-label transfer; the sign of the gain depends on the teacher–student capacity gap. On-device latency measured on a deployment-class server processor (single-input inference, 200 measurements) shows a 2.16× wall-clock speedup for the statically quantized 8-bit-integer SimpleCNN over its 32-bit floating-point counterpart, and the same 8-bit-integer binary builds, fits within 445 kilobytes of on-chip memory, and runs end-to-end on a simulated STM32H7 (Arm Cortex-M7) microcontroller target under the Renode hardware simulator. The toolkit is released under the MIT license.

Author 1: Mohammed Hassan Alnemari

Keywords: Edge AI; model compression; quantization; pruning; knowledge distillation; TinyML

Paper 76: RGB-D Bin-Picking System for Ergonomic Automotive Clip Assembly: 3D Annotation, Deep Learning, and 6-DOF Pose

Abstract: Automotive seat cover assembly entails the manual collection of plastic J-shaped clips from boxes located at a non-ergonomic distance from the sewing stations. Operators are compelled to inadvertently pick up multiple components at once, and make actions that involve repetitive stretching. The daily, consistent repetition leads to misassembly of the critical seat to frame connectors, and adds on to physical stress. This study showcases a fully integrated RGB-D bin-picking solution that uses depth-dependent grasp planning and deep learning object detection for interlocked plastic clip handling. GraspAnnotator Pro, a dual-modality bespoke software solution developed for this work, allows for model training on point cloud and RGB data for cluttered environment 6D pose estimation. This is achieved through a custom integrated annotation tool that simplifies the labeling of grasp pose and object boundary assignments. The system reduces strain on operators by automatically positioning parts within ergonomic zones and using fault-tolerant handling integrated with assembly verification. Real-world deployment validation over 6 weeks of continuous operation accumulated 3,000 pick-and-place cycles across 10 distinct J-shaped wire harness components, achieving a 93.7% first-attempt success rate with an average cycle time of 10.6 seconds. The system demonstrates a 42% reduction in cycle time compared to manual methods (18.3 seconds) with significant ergonomic improvements.

Author 1: Brahim Bergor Beguiel

Author 2: Ibrahim Hadj Baraka

Author 3: Yassir Zardoua

Keywords: Automated pick-and-place; RGB-D vision; YOLOv8 segmentation; point cloud matching; 6-DOF pose estimation; industrial robotics; automotive manufacturing; deep learning; quality inspection; PLC control

Paper 77: Urdu-ClimateGPT: Adapting LLM for Climate Data in Urdu Language

Abstract: Climate change is increasingly recognized as a major global challenge that affects environmental systems, weather patterns, and human societies around the world. Rising global temperatures have been linked to more frequent extreme weather events and long-term shifts in climate patterns. Communicating climate information effectively, therefore, becomes essential, especially in a way that is accessible and inclusive. But languages like Urdu are under-represented in the sources of climate knowledge, thus leaving many communities with fewer reliable sources of climate knowledge. To address this issue, Urdu-ClimateGPT is introduced by this study, as a domain-adapted language model based on LLaMA 3.1, along with a retrieval-augmented conversational framework built around it. Domain specific fine-tuning is combined with retrieval-based grounding evidence by the system. This is an effort to make the hallucinatory responses less common and factual disalignments in the generated responses less frequent, in the context of conversations with climate-related topics. The model was evaluated on a held-out set of Urdu climate prompts, and compared to the baseline LLaMA 3.1 model. The findings reveal that Urdu-ClimateGPT outperforms in various automated evaluation metrics including: language fluency, domain-specific correctness, factual consistency, and response completeness. Overall, a normalized average score of 0.82 was achieved by the Urdu-ClimateGPT, whereas a score of 0.52 was scored by the baseline model. These results suggest that large language models for low-resource languages in specialized domains can be adapted, which is both feasible and beneficial. It is shown by the study that hallucination like behavior can be reduced by retrieval augmented architectures when evaluated using automated metrics. However, further evaluation by human experts will be necessary to determine the system’s factual reliability and its potential real-world impact.

Author 1: Muhammad Farooq

Author 2: Muhammad Asif Habib

Author 3: Jabeen Sultana

Author 4: Muhammad Umar Aftab

Keywords: Climate change communication; domain-specific AI; large language models; low-resource languages; natural language processing; retrieval augmented generation; Urdu language processing

Paper 78: Leveraging MCESTA and Large Language Models for Next-Generation Similarity Learning

Abstract: Large Language Models (LLMs) have reshaped how machines read and compare text, yet most similarity learning pipelines built on top of them still behave like black boxes: a single cosine score is returned without any indication of why two sentences were deemed close. This study proposes a different route. We pair a fine-tuned Sentence-BERT (SBERT) encoder with MCESTA, a fuzzy multi-criteria aggregation layer that combines a semantic (cosine), a geometric (Manhattan) and a lexical (Jaccard) similarity through a small set of human-readable linguistic rules. The output is a single similarity score in [0, 1] that remains traceable to the rules that produced it: because the rule base contains only twelve Mamdani rules, the chain of fired antecedents can be inspected directly after each inference, which is what we mean by traceability in this study. We evaluate the framework on the Quora Question Pairs corpus (the public release contains about 404,000 question pairs in English with a 63/37 non-paraphrase to paraphrase split) against five strong baselines, including SimCSE and AnglE. Our model reaches Accuracy = 0.90 and AUC = 0.94, outperforming every baseline. A controlled ablation shows that the gains come from the fuzzy aggregation step itself, not from the choice of encoder, while a robustness study reveals that the soft membership functions absorb noise and threshold variations more gracefully than a plain cosine baseline. The fuzzy aggregation step runs in O(|R|) per pair, where |R| = 12 is the size of the rule base, so its computational overhead on top of the encoder forward pass is negligible. Adaptive fuzzy rules, multilingual similarity, and domain-specific deployments are positioned as future extensions rather than as results of the present study.

Author 1: Mohamedou Cheikh Tourad

Author 2: Abdelmounaim Abdali

Author 3: Mohamed Dhleima

Author 4: Naoual Mouhni

Author 5: Sana Chakri

Author 6: Ibtissam Amalou

Author 7: Saadbouh Cheikh El Mehdy

Keywords: Similarity learning; MCESTA; fuzzy aggregation; Natural Language Processing; Large Language Models; Sentence-BERT; explainable AI

Paper 79: Enhanced Pied Kingfisher Optimization Algorithm with Hovering Scouts and Foraging Flocks Mechanisms

Abstract: Population-based metaheuristic algorithms are widely used for solving nonlinear, nonconvex, constrained, and high-dimensional optimization problems. However, many swarm-based optimizers still suffer from premature convergence, loss of population diversity, and weak exploitation of multiple promising regions. To address these limitations, this study proposes the Hovering Scouts and Foraging Flocks Pied Kingfisher Optimizer (HSFFPKO), an enhanced variant of the Pied Kingfisher Optimizer (PKO). The proposed method introduces two complementary mechanisms. The Hovering Scouts mechanism applies scale-aware Gaussian probing with weak global-best guidance to restore local diversity and reduce stagnation, while the Foraging Flocks mechanism organizes the population into temporary subgroups guided by leader–centroid targets under a shrinking search radius. The performance of HSFFPKO was evaluated on the CEC 2017 benchmark suite and nine constrained engineering design problems. In the ablation study at D = 10, HSFFPKO achieved the best average rank of 1.67 and obtained 20 wins out of 30 benchmark functions. In the scalability analysis, HSFFPKO remained the best-ranked PKO variant at D = 30, D = 50, and D = 100, with average ranks of 1.37, 1.33, and 1.27, respectively. In the broad comparison with recent optimizers at D = 30, HSFFPKO obtained the best overall average rank of 2.60 and the highest number of function wins, with 11 wins out of 30 functions. The Nemenyi post-hoc test showed that HSFFPKO was statistically comparable with the strongest competitors, including BWSMA and GPSOM, while significantly outperforming several other methods. Engineering results further confirmed that HSFF-PKO is highly competitive for continuous constrained design problems, although its performance was weaker on the discrete gear-train problem. These results indicate that HSFFPKO is a scalable and competitive PKO variant for continuous numerical and engineering optimization.

Author 1: Aarhus Dela Cruz

Keywords: Exploration-exploitation balance; metaheuristic optimization; numerical optimization; Pied Kingfisher Optimizer; swarm intelligence

Paper 80: A Lightweight Explainable Hybrid Deep Learning Approach for Early Skin Disease Detection

Abstract: Early detection of skin pathology is critical for patient survival and treatment effectiveness, particularly in the case of aggressive malignancies such as melanoma. In this study, we propose a lightweight and explainable hybrid deep learning approach for early identification of skin diseases. The proposed approach combines a transformer-based module to capture the global contextual dependencies with a CNN for an efficient local feature extraction. In addition, Grad-CAM approach is used to increase the interpretability of the model and offer visual explanations for the forecasts. We evaluate the suggested technique on a publically accessible dermoscopic benchmark and achieve 93% accuracy, 92% precision, 91% recall and 92% F1-score, outperforming numerous mainstream designs. The experimental results imply that the proposed approach achieves a good trade-off between accuracy, computational economy, and interpretability for real-world and resource-limited medical applications.

Author 1: Mohammad Barr

Keywords: Medical image analysis; skin disease detection; explainable AI; hybrid model; Vision Transformer

Paper 81: Intrusion Detection System for Ransomware in Network Traffic Using Supervised Machine Learning

Abstract: Ransomware is one of the most dangerous cyber threats today, as it can disrupt systems and cause serious financial losses. Traditional detection methods often fail to catch newer attacks because they can hide within normal network traffic. In this study, we used machine learning to detect ransomware based on network data. We tested four models Logistic Regression, Decision Tree, Random Forest, and Gradient Boosting using the UGRansome dataset, both before and after balancing it with SMOTE. The Decision Tree model gave the best results, achieving 99.40% accuracy, 98.0% precision, 99.90% recall, and an AUC-ROC of 99.95%. We also found that protocol flags and network flow features played a key role in detecting attacks. Overall, using tree-based models with balanced data proved to be a simple and effective way to build a real-time ransomware detection system.

Author 1: Ziad Almulla

Author 2: Moath Alamri

Author 3: Mounir Frikha

Keywords: Ransomware detection; machine learning; IDS; network traffic analysis; real-time detection

Paper 82: Domain Knowledge-Enhanced Welding Robot Path Planning Algorithm

Abstract: Welding Robot Path Planning (WRPP) is a core technology for welding automation that seeks an optimal path that satisfies the requirements of the welding process in complex obstacle environments. However, existing algorithms generally lack the integration of knowledge in the welding domain and exhibit insufficient adaptability to complex working conditions. To address these issues, this study proposes a Domain Knowledge-Enhanced Welding Robot Path Planning Algorithm (DKE-WRPP) equipped with a pluggable knowledge fusion frame-work, which is validated on three representative path planning algorithms: Artificial Bee Colony (ABC), Genetic Algorithm (GA), and Optimal Rapidly-exploring Random Tree (RRT*). Specifically, we first employ Large Language Models (LLMs) to extract domain knowledge such as welding processes and safety distances, and generate standardized knowledge vectors via semantic encoding using a pre-trained language model. Then, a Knowledge Enhancement Module (KEM) is constructed to deeply fuse knowledge features and path geometric features through an attention mechanism, and adaptively update the cost functions of the three baseline algorithms, realizing low-intrusive coupling between domain knowledge and planning algorithms. Finally, experiments in a 300×300 grid environment demonstrate that, compared with traditional algorithms, the knowledge-enhanced algorithms reduce the convergence iterations by more than 33%on average and significantly improve path smoothness. The results fully verify the effectiveness of domain knowledge enhancement and the universality of the pluggable framework, providing an efficient and stable solution for welding robot path planning in complex working conditions.

Author 1: Ming Liu

Author 2: Peng Shao

Keywords: Path planning; intelligent optimization algorithm; welding robot; Large Language Models

Paper 83: Optimizing Whiteboard Digitization with Clarix: An Automated Vision-Based System for Real-Time Text Extraction and Surface Cleaning

Abstract: Whiteboard digitization in educational settings grapples with challenges posed by handwritten text, mathematical notation, and unstructured layouts that confound traditional Optical Character Recognition (OCR) systems. Clarix, a low-cost embedded system, addresses these issues by integrating a Raspberry Pi, high-resolution camera, and servo-driven erasure mechanism to automate content capture, text extraction, and archiving as searchable PDFs. This study benchmarks five multi-modal Large Language Models (LLMs) (GPT-4o, Claude, Gemini, DeepSeek, Grok) against traditional OCR systems (PaddleOCR, EasyOCR, TesseractOCR) for extracting whiteboard content across mathematics, physics, and economics domains. Results highlight a stark performance divide: multimodal LLMs achieved F1-scores ranging from 0.7550 to 0.8466, with Gemini leading at 0.8466 (precision 0.8188, recall 0.8784), followed by GPT- 4o (F1=0.8162) and DeepSeek (F1=0.7965), while PaddleOCR topped traditional systems with an F1-score of 0.3333 (precision 0.3523, recall 0.3223), followed by EasyOCR (0.1158) and Tesser-actOCR (0.0000). Notably, increased region detection correlated with diminished performance, underscoring the superiority of contextual understanding over exhaustive segmentation. Clarix’s fusion of intelligent automation and advanced text processing marks a transformative advancement in bridging analog and digital educational environments.

Author 1: Abdellah NABOU

Author 2: Ameksa Mohammed

Author 3: Ezzahoud Hajar

Author 4: Bazgour Yassine

Author 5: Abouhane Zahra

Keywords: Whiteboard digitization; multimodal Large language Models (LLMs); Optical Character Recognition

Paper 84: Adaptive Temporal Windowing for Streaming Outlier Detection Under Dynamic Arrival Rates

Abstract: Streaming outlier detection requires adaptive mechanisms capable of handling continuously evolving data streams under dynamic arrival rates. Existing count-based approaches fail under bursty and irregular stream arrival patterns commonly observed in real-world systems, since they trigger model updates after a fixed number of instances without considering temporal dynamics. In this study, we propose an adaptive time-driven windowing framework for streaming outlier detection that de-couples model updates from instance count and instead leverages elapsed time as the primary control mechanism. The proposed approach is based on the Density Incremental Local Outlier Factor (DILOF) and introduces a time-aware update strategy aligned with real-world streaming behavior. Extensive experiments on benchmark datasets demonstrate that the proposed method achieves robust and stable detection performance, with AUC values ranging up to 0.96. The results further show that time-based windowing provides a consistent trade-off between detection accuracy and computational efficiency, while offering a temporally grounded update mechanism for streams with variable arrival behavior. In addition, we analyze hybrid count-time strategies and demonstrate their limitations due to dominance effects. Repeated runs further indicate the robustness and consistency of the proposed framework. The findings highlight that temporal awareness is a critical factor in stream outlier detection and should be explicitly incorporated into windowing mechanisms, particularly in resource-constrained environments such as fog and edge computing.

Author 1: Hend Maher

Author 2: Mohamed Khafagy

Author 3: Heba Nagaty

Keywords: Streaming outlier detection; time-based windowing; Density Incremental Local Outlier Factor; data streams; anomaly detection; fog computing

Paper 85: Deep Learning-Based Road Damage Detection Using Improved YOLOv8: Model Performance and Implementation

Abstract: Road infrastructure monitoring plays a crucial role in ensuring safety and economic efficiency; however, conventional manual inspection methods are expensive and resource-intensive. This study presents the development and evaluation of a real-time road damage detection system using the improved YOLOv8 architecture. Building upon the strengths of previous YOLO models, YOLOv8 offers enhanced accuracy and inference speed, making it highly suitable for mobile deployment. The model was trained to identify six common road damage types in Indonesia: potholes, alligator cracks, transverse cracks, longitudinal cracks, edge cracks, and road joints. Utilizing a hybrid dataset of 2,946 images, combining locally collected data and the RDD2020 dataset, the system incorporates mosaic augmentation and optimized preprocessing to improve generalization. The optimized YOLOv8 model achieved a mean Average Precision (mAP@50) of 96.3%, an F1-score of 91%, and an overall accuracy of 91%, demonstrating superior detection and classification performance. The system was deployed as a user-friendly smartphone application, enabling automated, geo-tagged road condition surveys and offering road authorities a scalable, efficient, and practical tool for infrastructure monitoring.

Author 1: Aulia Rahman

Author 2: Hwa Jen Yap

Author 3: Rusdha Muharar

Keywords: Road damage detection; YOLOv8; deep learning; computer vision; mobile application

Paper 86: Smarter Bridges: Leveraging Artificial Intelligence to Reshape University-Industry Technology Transfer

Abstract: University-industry technology transfer (UITT) is essential for converting academic research into commercial use, yet traditional strategies often fail to address the knowledge gap. Literature suggests that institutional inertia, communication barriers, and ineffective marketing strategies hinder the commercialization of technology. This study proposes a conceptual framework that incorporates AI-driven marketing to enhance knowledge dissemination, market identification, and stakeholder engagement within the technology transfer process. This systematic literature review amalgamates insights from UITT, AI marketing applications, and knowledge management systems. A qualitative analysis of peer-reviewed literature from 2017 to 2025 identifies trends, deficiencies, and emerging patterns, leading to an integrated framework that assesses technology transfer strategies and the implementation of AI marketing across diverse sectors, leveraging the Technology-Organization-Environment (TOE) model and the Unified Theory of Acceptance and Use of Technology (UTAUT). The investigation demonstrates that AI-enhanced marketing can significantly bolster UITT through five AI-enhanced marketing capabilities: precise client segmentation, predictive analytics of market trends, tailored communication, improved knowledge management, and streamlined digital outreach. This methodology fosters reciprocal knowledge exchanges, positioning AI as a facilitator between market insights and university research aims while refining technology presentations for industry stakeholders. Moreover, the study highlights critical concerns regarding data privacy, implementation expenses, technical complexities, and the necessary proficiency in AI and technology transfer.

Author 1: Khaouja Mohammed

Author 2: Sanaa DFOUF

Author 3: Kaoutar ERRAKHA

Author 4: Hanan ELHARISSI

Author 5: Fekkak HAMDI

Keywords: Collaboration; university-industry; technology transfer; AI-enhanced marketing; innovation; knowledge sharing; economic growth; strategic partnerships; research initiatives; entrepreneurial mindset

Paper 87: Development of an Arabic Pet Adoption System with Hybrid Recommendations

Abstract: Pet adoption platforms often face challenges in effectively matching adopters with suitable pets due to limited personalization and the lack of localized language support. This study presents EWAA, an Arabic-enabled mobile platform for pet adoption in Saudi Arabia that integrates a hybrid recommender system combining content-based and collaborative filtering techniques. The content-based component constructs a user profile from one-hot encoded pet attributes (category, breed, color, and age) of previously liked pets, while the collaborative filtering component identifies similar users through cosine similarity and recommends pets based on their preferences. The two components operate independently in a mixed hybrid configuration, which mitigates the cold-start limitation of collaborative filtering and the over-specialization limitation of content-based filtering. The platform also emphasizes usability, accessibility, and privacy for both adopters and pet owners through a fully Arabic interface and controlled information visibility. The system was evaluated with 20 questionnaire participants and 6 interview participants using User Acceptance Testing (UAT) and Non-Functional Requirements (NFR) testing. Results indicate high levels of user satisfaction, with task completion rates of 100% on 10 of 11 test scenarios, page load times between 1 and 12 seconds, and learning times between 1 and 8 minutes, suggesting that the proposed approach provides a viable foundation for supporting the pet adoption process in Arabic-speaking contexts.

Author 1: Hailah Alballaa

Keywords: Recommender systems; pet adoption; mobile applications; collaborative filtering; content-based filtering; user experience

Paper 88: State-AttentNet: A Dynamic Volatility-Adaptive Hybrid Framework for Market State Classification in Frontier Economies

Abstract: The proper identification of latent market states is important for effective risk management. However, existing frameworks are often not good at distinguishing the difference between stable situations and systemic crashes in the high entropy environments of frontier economies. To address this challenge, this study presents State-AttentNet, an application-driven market-state classification framework for the Dhaka Stock Exchange. The framework combines a volatility-adaptive labeling scheme with a bidirectional LSTM encoder and temporal attention to classify stable and crash-like market states in the Dhaka Stock Exchange. The market sensitive technical features are synthesized with Bidirectional LSTM encoder of the temporal context into a hierarchical pipeline proposed methodology. This model is also supplemented by Adaptive Temporal Attention model to bring focus to high impact volatility events in a rolling window. Here, ‘crash’ means an adaptive volatility-stress state, not a return direction. Empirical evaluation on a longitudinal 26-year dataset shows that the model achieves 93% classification accuracy and an AUC of 0.97. It outperforms traditional baseline models in discriminating between crash and stable states. The results have significant practical implications for institutional investors as it is a trusted, automated detection instrument. This system is conducive for state contingent risk reduction measures and minimizing false alarms in instances of correction of incidental markets. Lastly, SHAP used to check the operational integrity of the model. This suggests that structurally informative lag signals and not stochastic noise determine classification decisions.

Author 1: Md. Abul Kalam Azad

Author 2: Abdul Kadar Muhammad Masum

Author 3: Najmus Saadat

Author 4: Esrat Jahan

Author 5: Ramona Birau

Author 6: Virgil Popescu

Author 7: Iuliana Carmen Barbacioru

Author 8: Stefan Margaritescu

Keywords: Frontier markets; hybrid deep learning; adaptive attention mechanism; dynamic volatility labeling; Explainable AI (XAI)

Paper 89: Cognitively Aligned Assessment Item Generation with Open-Source LLMs: A Comprehensive Evaluation on LearningQ

Abstract: Automated generation of high-quality educational assessment items is still difficult, especially when it comes to higher-order cognitive skills. Although Large Language Models (LLMs) show promise, their structural validity and cognitive alignment are limited. This study systematically evaluates fine-tuning open-source LLMs using an enriched LearningQ dataset that includes Bloom’s cognitive labels and evidence. The results show a clear performance contrast. Qwen2.5-3B-Instruct displays the best semantic reasoning, while Llama-3.2-3B shows better structural adherence, achieving a 94.9% validity rate and full compliance against answer leakage while maintaining high question validity. In contrast, older encoder-decoder models like FLAN-T5-XL do not generate valid questions. The study finds that small- to medium-sized instruction-tuned models, backed by strong data engineering, are successful at developing scalable, cognitively well-aligned assessment items.

Author 1: Mahmoud Badry

Author 2: Walaa Medhat

Author 3: Shereen A. Taie

Author 4: Asmaa Hashem Sweidan

Keywords: Automated question generation; Large Language Models; educational assessment; Bloom’s Taxonomy; Parameter-Efficient Fine-Tuning; LearningQ

Paper 90: A Machine Learning Approach for Targeting Lucrative Customer Segments

Abstract: Effective marketing plays a pivotal role in modern businesses, where strategic allocation of resources is essential for maximizing return on investment (ROI). Customer segmentation involves dividing users into distinct sub-groups based on common characteristics, enabling each segment to receive tailored promotions according to their behavior. However, identifying which customer segments to target can be financially challenging due to the significant costs associated with marketing campaigns. This study proposes a machine learning framework to forecast the profitability of various customer segments and optimize marketing strategies accordingly. The proposed solution leverages machine learning and deep learning techniques to classify customers based on their potential value. Specifically, conventional classifiers including AdaBoost, Gradient Boosting, Extreme Gradient Boosting (XGBoost), Linear Discriminant Analysis (LDA), Random Forest, Na¨ıve Bayes, Support Vector Machines (SVM), and K-Nearest Neighbors (KNN) are evaluated. In addition, deep learning models, namely 1D ResNet and 1D DenseNet, are investigated. All models are trained and evaluated under a unified protocol that includes SMOTE-based class im-balance handling, systematic hyperparameter tuning, and a joint analysis of predictive performance and computational cost. The experimental findings reveal that traditional models, particularly XGBoost and Gradient Boosting, consistently outperform deep learning models in terms of accuracy, precision, and computational efficiency, with both achieving the highest weighted F1 score of 0.62 while requiring nearly six orders of magnitude fewer computations than DenseNet-121. These results provide concrete evidence that ensemble tree-based methods are better suited than deep architectures for moderately sized, imbalanced, tabular marketing datasets.

Author 1: Leen Almajed

Author 2: Haifa Almakhdhoub

Author 3: Jana Alzeydan

Author 4: Alaa Bin Sleem

Author 5: Ouiem Bchir

Keywords: Return on investment; customer segmentation; machine learning; deep learning; profitability prediction; tabular data; Gradient Boosting; XGBoost

Paper 91: A Human-Centered Evaluation of AI-Generated Guidance: Integrated Statistical and Machine Learning Analysis with a Risk Framework for High-Stakes Domains

Abstract: The increasing use of large language models (LLMs) in domains requiring interpretation and judgment has raised critical questions about trust, reliability, and account-ability, particularly in contexts where decisions carry significant consequences. While prior work has focused primarily on improving system performance, limited attention has been given to how users evaluate and interact with AI-generated guidance in real-world, high-stakes settings. This paper addresses this gap through a large-scale empirical investigation of public perceptions of AI-generated religious guidance in Saudi Arabia. The analysis is based on survey data collected from 572 participants and combines quantitative statistical methods with a machine learning-based pipeline for analyzing open-ended responses. The quantitative component examines patterns in trust, perceived risk, privacy concerns, credibility, and user practices, while the qualitative component employs embedding-based clustering using Bidirectional Encoder Representations from Transformers (BERT), Uniform Manifold Approximation and Projection (UMAP), and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), followed by expert interpretation to derive structured parameters. The results indicate a cautious and conditional engagement with AI systems, characterized by moderate usage, low levels of trust, and strong concerns regarding reliability and source credibility. Users frequently verify AI-generated outputs and demonstrate a preference for human expert validation, particularly in complex or sensitive cases. Building on these insights, the study introduces a layered taxonomy of perceived risks spanning epistemic, reasoning, interactional, and institutional dimensions, providing a structured analytical framework for understanding how technical limitations translate into broader behavioural and governance challenges. These results highlight the importance of aligning AI system design with user expectations, emphasizing transparency, verifiability, and human oversight. The proposed taxonomy and analytical framework provide a foundation for future research and contribute to the development of governance approaches for AI systems deployed in high-stakes interpretive domains.

Author 1: Omar Al-Turki

Author 2: Felwah Alqahtani

Author 3: Eman Alqahtani

Author 4: Sarah Alswedani

Author 5: Sami Alshmrany

Author 6: Rashid Mehmood

Keywords: Large language models; AI-generated guidance; user perception; trust in AI; perceived risk; source credibility; human-AI interaction; high-stakes domains; risk taxonomy; risk framework

Paper 92: D-HAN-Net: A Hybrid Dual-Stream Architecture for Corporate Bankruptcy Prediction via Multimodal Fusion

Abstract: Early detection of corporate bankruptcy is essential for maintaining financial market stability and reducing systemic risk. However, existing predictive models often struggle under conditions of extreme class imbalance. Traditional approaches either analyze financial ratios or textual disclosures in isolation, while conventional cross-modal fusion strategies tend to dilute rare distress signals when integrating structured accounting metrics with qualitative management narratives. Consequently, subtle indicators of impending corporate failure are frequently overshadowed by dominant non-distress patterns, limiting the effectiveness of existing predictive systems. To bridge this methodological limitation, this study proposes D-HAN-Net, a hybrid dual-stream deep learning architecture. This framework is in-tended to address minority class suppression by dynamically balancing structured financial ratios with unstructured textual disclosures. Specifically, the model processes numerical indicators through a gated residual network to capture complex patterns. Simultaneously, it extracts semantic cues from corporate reports via a FinBERT-driven bidirectional GRU. These dual modalities are then aligned using a learnable cross-modal attention fusion gate, jointly optimized with focal loss. Experimental evaluations on a comprehensive multimodal dataset, utilizing stratified splits demonstrate that D-HAN-Net significantly outperforms state-of-the-art baselines, achieving a predictive accuracy of 94.00%, an F1-score of 88.00%, and an AUC of 0.9734. Practically, this framework equips investors, financial institutions, and regulatory authorities with a decisive early warning system. It enables proactive risk management by detecting subtle distress signals before corporate failure becomes irreversible. Furthermore, extensive stability testing and ablation analysis confirm that the model’s superior predictive reliability is highly robust against sampling uncertainty, fundamentally relying on the synergistic integration of all its architectural modules.

Author 1: Md. Abul Kalam Azad

Author 2: Abdul Kadar Muhammad Masum

Author 3: Najmus Saadat

Author 4: Esrat Jahan

Author 5: Ramona Birau

Author 6: Virgil Popescu

Author 7: Iuliana Carmen Barbacioru

Author 8: Stefan Margaritescu

Keywords: Bankruptcy prediction; multimodal deep learning; corporate disclosures; cross-modal fusion; class imbalance; early warning system

Paper 93: A Reliability-Aware Visual-Inertial Odometry for Dynamic and Low-Texture Environments

Abstract: Visual-inertial odometry (VIO) tends to degrade in aggressive dynamic and low-texture environments, where rapid motion, weak visual structure, and moving objects reduce the reliability of visual observations. This study presents TRAIL-VIO, a temporal reliability-aware visual-inertial odometry framework with line feature enhancement. The method estimates temporal observation reliability by combining semantic priors with IMU-based motion consistency, which allows a continuous and time-varying assessment of observation quality instead of frame-wise decisions. A reliability-aware point-line association scheme is also introduced, where inertial prediction is used to constrain feature matching and partially corrupted line segments are selectively retained. In addition, a reliability-guided marginalization strategy is applied to reduce the influence of unreliable visual constraints before they are incorporated into the prior. Experiments on the EuRoC MAV benchmark and a self-collected UAV dataset show that TRAIL-VIO achieves average RMSE values of 0.042 m and 9.51 m, respectively, outperforming representative baseline methods in dynamic and low-texture scenarios. Additional ablation, parameter-sensitivity, and runtime analysis further verify the contribution of the main modules, the robustness of the selected parameters, and the computational feasibility of the proposed framework.

Author 1: Yelu Liu

Author 2: Ruokun Qu

Author 3: Mengcheng Xu

Author 4: Chenglong Li

Author 5: Hui Jiang

Keywords: Visual-inertial odometry; point-line feature fusion; dynamic environments; observation reliability estimation; marginalization; UAV navigation

Paper 94: Trustworthy App Detection in Saudi Mobile Finance: Bridging Deep Learning and Interpretability

Abstract: Understanding user trust in mobile financial applications is crucial as these platforms increasingly shape how users in Saudi Arabia manage finances and engage with digital banking. However, existing deep learning–based detection models often function as black boxes, offering limited interpretability, while traditional machine learning models, though more transparent, fail to capture complex interactions between permissions, reviews, and user behaviors. To address this gap, we propose Trustworthy App Detection for Saudi Arabia (TAD-Saudi) - a novel framework for interpretable and behavior-aware trust evaluation. The frame-work integrates the representational power of deep learning with the explainability of simpler models, enabling both global and local interpretation of trust-related features. Experimental results show that TAD-Saudi outperforms traditional baselines across multiple models. Moreover, the analysis reveals that users may continue to trust applications requesting sensitive permissions, particularly when these apps have high ratings or positive reviews.

Author 1: Raed Alharbi

Author 2: Maryam Alghamdi

Keywords: XAI; trustworthy apps detection; encoding

Paper 95: Benchmarking Deep Learning for PM2.5 Forecasting in IoT-Based Smart Cities: GCN Spatial Encoding and Transformer Temporal Modeling

Abstract: Graph convolutional networks are widely used in air quality forecasting, yet their benefit over simpler approaches remains insufficiently validated. This study presents a fully reproducible benchmark comparing five model families (ARIMA, XG-Boost, LSTM, CNN-LSTM, and a GCN+Transformer model) on the public Beijing Multi-site Air Quality Dataset. All experiments share identical preprocessing, three random seeds (confidence intervals reported as exploratory), and a chronological split, with full code availability. ARIMA outperforms deep learning on aggregated data, due to strong temporal autocorrelation. On individual stations, Transformer-based models achieve the best performance through improved temporal modeling. A systematic multi-node analysis reveals that graph-based spatial aggregation can degrade performance under highly homogeneous conditions: when monitoring stations exhibit strong correlations, a simple linear baseline outperforms the GCN topologies tested here. Alternative spatial encoders (e.g., GAT, learnable adjacency) may behave differently and remain an open question. These findings define a practical regime in which GCN-based spatial aggregation provides no benefit over linear baselines. A preliminary compression experiment on Apple M1 reports a 42% inference-latency reduction via 8-bit quantization; this is indicative only, and validation on Raspberry Pi 4 or NVIDIA Jetson Nano is identified as required follow-up.

Author 1: Abdessamad BADOUCH

Author 2: Kaoutar BELHOUCINE

Keywords: Air quality forecasting; graph neural networks; IoT smart cities; transformer; PM2.5 prediction

Paper 96: A Controlled Benchmark of Open-Source and Proprietary LLMs for Few-Shot Microbiome IBD Classification

Abstract: Phenotyping inflammatory bowel disease (IBD) from gut microbiome profiles remains challenging due to 93% genus-level zero-inflation, skewed amplicon count distributions, and the practical cost of assembling labelled cohorts. Few-shot in-context learning (ICL) with large language models (LLMs) sidesteps the annotation bottleneck, yet existing benchmarks test only proprietary APIs—a deployment model incompatible with the data governance constraints of most clinical sites. We benchmark six frontier LLMs under identical few-shot conditions—three proprietary (GPT-4o, Claude claude-opus-4-5, Gemini 2.5 Flash Lite) and three open-source (Mistral 7B, LLaMA-3 8B, DeepSeek-R1-Distill-Qwen 1.5B)—evaluated against supervised baselines (Random Forest, XGBoost, LightGBM, soft-voting Ensemble) on 16S rRNA amplicon data (n = 1,316; holdout n = 30). All six LLMs received identical prompts, the same log-normalised top-20 features, and the same random shot selection strategy with a fixed seed, ensuring that observed differences are attributable to model capacity rather than experimental conditions. The supervised Ensemble led holdout performance (Macro-F1: 0.7948; AUC: 0.7725). Among LLMs, Mistral 7B achieved the highest Macro-F1 (0.5417), surpassing all three proprietary models without any parameter update. The mean performance gap between open-source (0.4711) and proprietary (0.5101) groups was only 0.039 Macro-F1 points—too narrow to justify exclusive reliance on commercial APIs in privacy-sensitive deployments. Within-open-source variance (0.130 Macro-F1 points) substantially exceeded this inter-family gap, indicating that model selection within the open-source ecosystem is the more consequential practical decision. These results suggest that open-source LLMs running on local hardware are a workable option when labeled data is limited and routing patient-derived data to external APIs is not permitted.

Author 1: Nouhaila En Najih

Author 2: Soufiane Hamida

Author 3: Ahmed Moussa

Keywords: Inflammatory bowel disease; gut microbiome; few-shot learning; large language models; open-source LLMs; pro-prietary LLMs; in-context learning; 16S rRNA; log-normalisation; zero-inflation; GPT-4o; Claude; Gemini; LLaMA; Mistral; DeepSeek

Paper 97: A Lightweight and Robust APBT–LBP Deep Feature Zero-Watermarking Framework with DNA Encryption for Medical Images

Abstract: The protection of medical images in healthcare services has become important due to the adoption of telemedicine and cloud-based healthcare services. Conventional watermarking methods embed information directly into the host image, which may introduce subtle distortions and potentially affect diagnostic accuracy. This paper addresses the problem by introducing a robust zero-watermarking scheme that preserves the original medical image without any modification. The proposed approach is designed on a hybrid feature extraction scheme. Initially, desired regions of interest are identified using variance-based analysis. Local Binary Pattern (LBP) is then applied to capture fine texture details. Further, the All-Phase Biorthogonal Transform (APBT) is used to obtain stable low-frequency in-formation. These features are subsequently pipelined through the lightweight VGG16 convolutional neural network to extract high-level semantic representations. The resulting features are fused, normalized, encrypted, and converted into binary form using Quantization Index Modulation (QIM). DNA encryption is applied to the watermark to produce a secure zero-watermark key that is stored externally. Extensive experiments conducted under a wide range of signal processing and geometric attacks show the effectiveness of the proposed method. The results show high normalized correlation above 0.99, bit error rates (≤ 0.1%), and complete preservation of higher image quality, making the approach suitable for medical image authentication and copyright protection.

Author 1: Ranjan Kumar Senapati

Author 2: Prasanth Mankar

Author 3: B Padmaja

Author 4: Chilamakuru Nagesh

Author 5: Pradeep Kumar

Author 6: Gandikota Ramu

Author 7: Gandharba Swain

Keywords: Zero watermarking; medical image authentication; deep learning; APBT transform; local binary pattern; CNN feature extraction; robust watermarking

Paper 98: A Privacy-Aware Federated Hybrid Model for Multimodal Mental Health Analysis

Abstract: As mental health disorders such as stress, anxiety, depression, and post-traumatic stress disorder (PTSD) affect a substantial part of the world population, current diagnostic methodologies are still centralized, subjective, and sensitive to privacy concerns. To mitigate these limitations, this study presents a new framework for multimodal mental health classification within a privacy-preserving federated learning framework learning using electroencephalography (EEG), electrocardiography (ECG) and galvanic skin response (GSR) signals. Furthermore, we pro-pose a hybrid deep learning architecture, which combines CNN-LSTM-Transformer blocks to effectively learn spatial, temporal, and long-range dependencies within physiological signals. After preprocessing the cleaned data through artifact removal, band-pass filtering, normalization and multimodal feature fusion signal quality is improved. The proposed model is trained in a federated setting with multiple clients for decentralized training without sharing raw data allowing it to preserve privacy and communication efficiency supporting non-IID data extensions. We evaluate on two datasets, SAM40 (stress detection) and DAPS (anxiety, depression, and PTSD classification). The proposed framework achieved 97% accuracy on SAM40 and more than 96% accuracy on DAPS. Comparative assessments with recent federated and centralized methods validate its strength in multimodal fusion and robust feature exploitation. These results demonstrate the possibility of a general framework for designing privacy-preserving and efficient mental health monitoring systems that can support both Clinical and Wearable-device applications.

Author 1: Yusra

Author 2: Riaz UlAmin

Keywords: Mental health detection; hybrid CNN-LSTM-transformer; federated learning; multimodal physiological signals

Paper 99: Design and Evaluation of a Ubiquitous Learning Game for Adolescent E-Cigarette Addiction Prevention

Abstract: This study presents the design, development, and pilot evaluation of Addict-Shield, a mobile educational serious game aimed at preventing electronic cigarette use among adolescents. The proposed approach is grounded in the DICE model (Define, Imagine, Create, Evaluate) to ensure a structured and learner-centered design process. The game integrates narrative-driven scenarios, gamification mechanics, and scientifically validated prevention content targeting risk awareness, stress management, and refusal skills. A pilot pre-test/post-test study was conducted with a convenience sample of 25 adolescents aged 12–18 years. Results indicate substantial improvements in knowledge scores (mean increase from 32% to 80%), stronger negative attitudes toward vaping, and increased behavioral intention to refuse e-cigarettes in peer-pressure situations (from 33% to 85%). Descriptive comparisons between pre-test and post-test results indicated substantial improvements in knowledge, attitudes, and behavioral intention following the intervention. Participants also reported high engagement and perceived usefulness (83–91%). Although the findings remain preliminary due to the limited sample size and absence of a control group, the results suggest the potential of structured mobile serious games for adolescent addiction prevention. These findings highlight the relevance of mobile serious games as innovative tools for adolescent addiction prevention and health education.

Author 1: Lamyae Bennis

Author 2: Khalid Kandali

Author 3: Hamid Bennis

Keywords: Serious games; mobile learning; adolescent health; vaping prevention; gamification; DICE model; pilot study

Paper 100: MILP for Multimodal Urban Transport: Formal and Informal Sectors

Abstract: This study presents a mixed-integer linear programming (MILP) formulation for optimizing multimodal urban transportation networks that integrate formal and informal public transport sectors. Whereas prior optimization approaches have predominantly addressed regulated formal services, the proposed model explicitly incorporates informal operators prevalent in many African cities. The objective is to minimize total user travel time, comprising in-vehicle time, waiting time at stops, and trans-fer time between modes, subject to flow conservation constraints, limits on mode changes, and sector-specific operational rules. Under the assumption of static demand and known arc travel times, computational experiments use realistic synthetic instances calibrated to reflect operating characteristics of the operators providing public transport services in Abidjan, Cˆote d’Ivoire—informed by close observation of non-public operational patterns and published mobility statistics—with 5 to 24 logical stops and up to 2,487 active arcs. Compared with formal-only system con-figurations, the integrated formal–informal formulation reduces total travel time by up to 53% on the largest tested instances (24 stops; with smaller gains on smaller networks). All instances were solved with IBM CPLEX 22.11; maximum solve times were 1.63 s on corridor instances (n ≤ 24) and 0.69 s on an extended set with n ≤ 100 (Intel Core i7, 16 GB RAM). These findings indicate that coordinated planning of formal and informal transport can materially improve urban mobility in developing cities.

Author 1: Oumar KONE

Author 2: Yapi Fiacre Aristide EDI

Author 3: Kouassi Hilaire EDI

Author 4: Pawoumodom Matthias TAKOUDA

Keywords: Mixed-integer linear programming; multimodal transportation; urban mobility; formal transport; informal transport; operations research; network optimization

Paper 101: Artificial Intelligence in Motorsport: A Scoping Review of Applications, Challenges, and Research Gaps

Abstract: This study presents a state-of-the-art review of Artificial Intelligence and Machine Learning applications in motorsport, with a particular focus on Formula 1. As modern racing generates increasingly large volumes of high-frequency telemetry data, traditional physics-based analysis methods have proven insufficient to capture the complex, non-linear relation-ships between vehicle, driver, and environment. AI has emerged as a critical tool for extracting actionable insights from this data across six core domains: vehicle performance optimization, race strategy, autonomous racing and real-time decision support, telemetry and race data analysis, driver coaching and simulation, and predictive maintenance. This review further examines transferable methodologies from adjacent fields — particularly autonomous driving research — and assesses their relevance to high-performance motorsport contexts. A systematic analysis of peer-reviewed publications and preprints reveals a highly uneven distribution of research, with race strategy and vehicle optimization relatively well-covered, while predictive maintenance and computer vision applications remain largely unexplored.

Author 1: Radu C. Lucaciu

Author 2: Cornelia Aurora Gyorödi

Keywords: Artificial Intelligence; Machine Learning; motor-sport; Formula 1; scoping review; PRISMA-ScR; race strategy optimization; telemetry analysis

Paper 102: Benchmarking Deep Vision Architectures: From Controlled Datasets to Real-World Clinical Validation

Abstract: Stroke is a leading cause of mortality and long-term disability, making rapid and reliable detection from non-contrast computed tomography (CT) scans essential for timely clinical intervention. This study introduces NeuroCT-Bench, a unified and reproducible benchmark for evaluating deep learning architectures for automated stroke classification from brain CT images. The benchmark systematically compares five Convolutional Neural Networks (VGG16, ResNet50, DenseNet121, EfficientNetB0, MobileNetV2) and four Vision Transformers (Swin Transformer, ViT, DeiT, PVT-Small) under identical preprocessing, augmentation, and evaluation protocols using the Brain Stroke CT Image dataset from Kaggle, comprising 1,551 normal and 950 stroke slices. Internal validation using a deterministic 80/20 stratified split (seed = 42) demonstrated near-perfect performance for Transformer-based models, with PVT-Small and Swin Transformer achieving ROC-AUCs of 99.96% and 99.98%, respectively, while DenseNet121, VGG16, and EfficientNetB0 achieved strong CNN baselines with ROC-AUCs of 99.93%, 99.67%, and 99.05%. To evaluate robustness under real-world domain shift, the top-performing models were further assessed on an external patient-level clinical dataset containing 530 CT studies. EfficientNetB0 demonstrated the strongest generalization capability (accuracy: 76.92%, precision: 83.85%, ROC-AUC = 88.0%), whereas high-capacity Transformer models exhibited substantially larger performance degradation (ROC-AUCs of 75.0% for PVT-Small and 67.0% for Swin Transformer). These findings highlight the discrepancy between curated public datasets and heterogeneous clinical imaging conditions, emphasizing that high internal performance does not necessarily guarantee clinical robustness. In addition, an ablation study was conducted to evaluate a lightweight CNN–Transformer gated fusion strategy. Results demonstrate that adaptive fusion improves robustness and generalization compared with individual CNN or Transformer models and static feature concatenation. Overall, NeuroCT-Bench provides a transparent and reproducible framework for evaluating deep learning models for stroke analysis and supports future development of clinically deployable hybrid CNN–Transformer systems.

Author 1: Raghda Essam Ali

Author 2: Noha Ahmed Saad El-Dien

Author 3: Magi Hossam Eldin Mahfouz

Keywords: Stroke detection; brain CT image; deep learning; Convolutional Neural Networks; Vision Transformers; clinical validation; neuroimaging AI

Paper 103: Smart IoT Framework for Sustainable Cities: Integrating Information Systems and Artificial Intelligence

Abstract: IoT architecture is an intelligent green city when it incorporates information systems and Artificial Intelligence (AI) to maximize the utilization of resources, upgrade public services, and produce cities very efficient but also environment-friendly. There remain poor feature selection and extraction, unbalanced load, Quality of Service (QoS) issues, and energy consumption by blockchain which prevents effective implementation. This research solves these issues through a multi-step approach. First, we build a smart IoT network that gets the IoT device data from our CICIDS2017 dataset, preprocessed with Normalization to standardize the data. Next, Wrapper-based Feature Selection is applied to identify the most significant features used by the Autoencoder's deeper Feature Extraction to improve model performance. A QoS scheme based on Software Defined Networking (SDN) will dynamically distribute loads with low latency to balance loads and achieve QoS. We further utilize a stateless Q-learning algorithm to avoid congestion in IoT device data distribution. We then use the Hyperledger Fabric for efficient blockchain in combination with Linear Network Coding (LNC) to save energy on the system. To detect cyber-attacks, we adopt a Quasi-Recurrent Neural Network (QRNN) that is Bio-optimized to enhance detection while minimizing false positive responses. Finally, we evaluate the performance of the proposed system on the following metrics: Response Time (ms), Throughput (Mbps), Attack Detection (%), False Positive Rate (%), and Energy Consumption (%). This approach is implemented through NS-3.35 using Python, providing a solid framework for comparative studies and the promotion of sustainable operations of a smart city.

Author 1: Amarildo Rista

Author 2: Alma Stana

Keywords: Smart city; blockchain; IoT; artificial intelligence; auto encoder; Quality of Service