IJACSA Volume 16 Issue 10 - thesai.org

IJACSA Volume 16 Issue 10

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: AI-Assisted Workflow Optimization and Automation in the Compliance Technology Field

Abstract: Against the backdrop of digital transformation and stricter regulation, enterprise compliance work demands higher efficiency and accuracy. The auxiliary compliance process has become an important entry point for optimizing the compliance system due to its strong transactional nature and high degree of repetition. This study focuses on the process characteristics of auxiliary compliance work, sorts out its structural composition and organizational mechanism, proposes an optimization path with process reengineering, system modeling, and technology integration as the core, and focuses on exploring the collaborative application of key technologies such as RPA, rule engine, and semantic recognition in process automation. Research suggests that the systematic optimization and intelligent upgrading of auxiliary processes will help build a modern compliance operation system that is responsive, efficient, structurally clear, and risk controllable.

Author 1: Zhen Zhong

Keywords: Compliance technology; auxiliary process; process optimization; automation

Paper 2: From Legacy to Cloud: Migration Strategies for Traditional Financial Institutions Using AWS

Abstract: Traditional financial institutions face unprecedented pressure to modernize their technological infrastructure while maintaining regulatory compliance and operational stability. This research examines the strategic approaches, implementation challenges, and outcomes of migrating legacy banking systems to Amazon Web Services (AWS) cloud infrastructure through a mixed-methods analysis of twelve financial institutions that completed migrations between 2019 and 2024. Through structured interviews with technology leaders and quantitative analysis of migration outcomes, including regulatory considerations and real-world implementation cases, this study identifies key success factors and potential pitfalls in large-scale financial services cloud adoption. The research reveals that institutions adopting phased migration strategies with robust risk management frameworks achieve 92% success rates with 30-45% cost reductions and 40-60% performance improvements, compared to 58% success rates for rapid, wholesale transitions. Furthermore, the study demonstrates that AWS-specific services such as AWS Control Tower and AWS Config provide essential governance capabilities that traditional financial institutions require for regulatory compliance during cloud transformation initiatives.

Author 1: Uday Kiran Chilakalapalli

Author 2: Brij Mohan

Author 3: Vinodkumar Reddy Surasani

Keywords: Cloud migration; financial services; AWS; compliance; risk management; governance

Paper 3: From Logs to Knowledge: LLM-Powered Dynamic Knowledge Graphs for Real-Time Cloud Observability

Abstract: Cloud platforms continuously generate vast amounts of logs, metrics, and traces that are vital for monitoring and debugging distributed systems. However, current observability solutions are often siloed, dashboard-centric, and limited to surface-level correlations, making it difficult to derive actionable insights in real time. In this work, we present Log2Graph, a novel framework that leverages large language models (LLMs) to transform heterogeneous telemetry into dynamic knowledge graphs that evolve alongside system state. Unlike traditional log analytics, Log2Graph unifies unstructured messages, distributed traces, and configuration data into a living graph representation, enabling real-time dependency mapping, causal chain analysis, and compliance monitoring. Furthermore, the framework supports natural language queries over the evolving graph, allowing operators to ask questions such as “what services will be impacted if this database fails?” and receive precise, graph-backed explanations. Our evaluation on multi-cloud testbeds shows that Log2Graph reduces incident resolution time, improves accuracy in dependency detection, and enhances operator productivity. This work introduces a new paradigm of LLM-augmented observability, bridging the gap between raw logs and actionable cloud intelligence.

Author 1: Nurmyrat Amanmadov

Author 2: Tarlan Abdullayev

Keywords: Large Language Models (LLMs); AI for cloud computing; knowledge graphs; logs

Paper 4: IoT-Enabled Data-Driven Optimization of Dynamic Thermal Loads for Low-Energy Buildings

Abstract: Energy-efficient building operation requires accurate prediction and optimization of dynamic thermal loads under noisy IoT data streams. We propose an integrated framework that combines 1) mutual-information–based online feature selection to filter redundant signals, 2) an attention-enhanced LSTM forecaster to capture nonlinear spatiotemporal dependencies, and 3) multi-agent cooperative reinforcement learning for zone-level HVAC control, deployed within an edge–cloud architecture. Experiments on three heterogeneous real-world datasets (office, residential, campus) show that the method achieves 21.7% median energy savings (IQR 18.9–23.1%), improving over MADDPG by +5.8 percentage points (p=0.004, Wilcoxon). Forecasting accuracy is also improved, with MAE reduced by 16.7% (95% CI 12.4–20.1%) compared with Seq2Seq+Attention. Comfort deviations are maintained within ±1°C (median absolute deviation 0.32°C). Robustness tests indicate graceful degradation under σ≤0.2 Gaussian noise and ≤20% missing data, while ablation confirms the contribution of each module. Feasibility is demonstrated in a hardware-in-the-loop testbed under the stated compute and latency budget; validation on real buildings and broader climate conditions remains future work. This study contributes to smart building energy management, IoT-based HVAC control, and sustainable operation optimization.

Author 1: Zhaojiang Lyu

Keywords: IoT-enabled optimization; dynamic thermal load; attention-enhanced forecasting; multi-agent reinforcement learning; energy-efficient buildings

Paper 5: Quantifying Career Preferences and Perceptions of Software Testing Among Filipino IT Students: A Mixed-Method Analysis

Abstract: Software testing (ST) careers have consistently demonstrated low appeal among IT students globally, creating significant workforce gaps in this essential field of information technology. This study investigates the extent to which Filipino IT students share this disinterest in software testing careers as observed in previous international studies, while examining the unique cultural, economic, and curricular realities that influence their career decision-making processes. Utilizing a mixed-methods approach, the research analyzes quantitative survey responses and qualitative focus group discussions to determine student perceptions and attitudes toward the software testing profession. The study employs a multidimensional framework to explore local factors that shape career preferences among Filipino IT students. Findings confirm that software testing is not the first career choice for most respondents, paralleling previous international research findings. However, qualitative data reveal that students demonstrate significantly greater interest when opportunities offer competitive salaries, clear career growth trajectories, meaningful professional development opportunities, and comprehensive academic training in software testing methodologies. The research identifies unique local factors, including economic pressures, cultural perceptions of professional prestige, and significant curriculum gaps that systematically influence students' career decisions. These results highlight critical needs for effective reforms within current IT curricula and enhanced career guidance programs to address the systematic undervaluation of the software testing profession. The study's implications suggest that targeted educational interventions and improved industry-academia collaboration could better prepare students for the fast-evolving demands of the IT industry while addressing the persistent shortage of qualified software testing professionals in both local and global markets.

Author 1: Chrisza Joy M. Carrido

Author 2: Abeer Alsadoon

Author 3: Thair Al-Dala’in

Author 4: Ahmed Hamza Osman

Author 5: Abubakar Elsafi

Author 6: Azhari Qismallah

Author 7: Albaraa Abuobieda

Keywords: Software testing; career preferences; Filipino IT students; mixed methods; education reform

Paper 6: Pedestrian Navigation System with 3D Map and Charging Server Based on Steganography

Abstract: A pedestrian navigation system with a steganography-based 3D map and billing server is proposed. The proposed system includes a server system that provides topographical maps and navigation information to both pedestrians and vehicles. When using the proposed system, necessary images of cross sections, intersections, or points of interest can be automatically obtained, similar to the Street View feature in Google Maps. Users can post photos taken with their camera phones and earn points if the photos are marked as posted. While the proposed system incurs a usage fee, these points can be used to reduce the subscription fee. If the quality of an image is superior to that of a previously archived image, the new image overwrites the previous one. The billing system for the system usage fee incorporates digital steganography for security reasons. This prevents the leakage of user information and other data. Through experiments, it is confirmed that the proposed system works well.

Author 1: Kohei Arai

Keywords: Pedestrian navigation; steganography; client-server system; geographic information; system; GIS

Paper 7: Comparative Evaluation of CNN Architectures for Skin Cancer Classification

Abstract: Skin cancer is one of the fastest-growing health problems worldwide. Early and accurate diagnosis is essential for improving treatment success and patient survival. However, many previous studies have focused on single CNN architectures or limited datasets, resulting in models with restricted generalizability. To address this gap, this study presents a comparative evaluation of three deep learning architectures (DenseNet169, MobileNetV2, and VGG19) for automatic classification of benign and malignant skin cancers using dermoscopic digital images. A total of 10,000 images were compiled from three public Kaggle datasets, preprocessed through resizing and data augmentation, and trained using transfer learning based on ImageNet weights. Two data split schemes (60:20:20 and 80:10:10) were applied to assess model robustness. Experimental results show that DenseNet169 achieved the highest test accuracy of 90.7 per cent, while MobileNetV2 was the fastest with an inference time of 16 seconds. These findings highlight the tradeoff between accuracy and computational efficiency and support the use of deep learning models, particularly DenseNet169 and MobileNetV2, in the development of real-time AI-assisted skin cancer diagnostic systems.

Author 1: Taopik Hidayat

Author 2: Nurul Khasanah

Author 3: Elly Firasari

Author 4: Laela Kurniawati

Author 5: Eni Heni Hermaliani

Keywords: Artificial intelligence; convolutional neural network; deep learning; dermoscopic images; skin cancer classification

Paper 8: TabNet–XGBoost Hybrid Model for Student Performance Prediction and Customized Feedback

Abstract: Virtual Learning Environments (VLEs) have emerged as a cornerstone of modern education, enabling large-scale delivery of learning materials, assessments, and interactions in fully or partially online formats. The dynamic and self-paced nature of VLEs makes the early prediction of learner scores crucial for timely intervention and support. The existing frameworks either underperform in capturing complex, non-linear relationships in heterogeneous educational data or lack interpretability mechanisms necessary for actionable interventions. This study proposes a TabNet–XGBoost hybrid model with SHAP-based interpretability for score range classification in VLE contexts, using the Open University Learning Analytics Dataset (OULAD). Data preprocessing involved cleaning, encoding, normalization, feature engineering, and score band derivation, producing an enriched feature matrix integrating demographic, assessment, and engagement indicators. TabNet’s sequential attentive feature selection extracted a latent representation of the most informative variables, which was subsequently refined by XGBoost to produce sharper decision boundaries for four distinct score ranges. SHAP values were computed post-prediction to identify domain-specific performance drivers, enabling alignment with a structured feedback module across seven predefined learning domains. Experimental results demonstrated a classification accuracy of 98.8% on the test set, outperforming the baseline frameworks. The SHAP-driven feedback mechanism provided interpretable, domain-targeted insights, enhancing the model’s practical applicability for educators and academic support teams. By integrating high predictive accuracy with transparent reasoning and actionable feedback, the proposed framework addresses both the technical and pedagogical requirements of early performance prediction in online learning environments, offering a scalable solution for real-time academic monitoring and intervention.

Author 1: Anupama Prasanth

Keywords: Virtual learning environments; student performance prediction; TabNet; XGBoost; SHAP; feedback generation; quality education

Paper 9: Enhanced Fault Detection in Software Using an Adaptive Neural Algorithm

Abstract: Software fault detection is crucial for ensuring reliable and high-quality software systems. However, traditional fault detection methods often rely on manual inspection or rule-based techniques, which are time-consuming and prone to human errors. In this research, the researchers propose an enhanced fault detection approach using an adaptive neural transfer learning algorithm. The goal is to leverage the power of neural networks and adaptability to improve fault detection accuracy and classification performance. The problem addressed in this research is the need for more effective fault detection methods that can handle the complexities of modern software systems. Existing fault detection techniques lack adaptability and struggle to cope with diverse software scenarios. Neural networks have shown promise in pattern recognition and classification tasks, making them suitable for fault detection. However, fixed architectures and training strategy limit their performance in different software contexts. To address this problem, the research proposes an adaptive neural transfer learning algorithm for fault detection. The algorithm dynamically adjusts its neural network architecture and training process based on the characteristics of the software under test. It incorporates adaptive mechanisms, such as adjusting learning rates and regularization techniques, to optimize performance. Real-time feedback and performance evaluation during the training process drive the adaptive mechanisms. To evaluate the proposed approach, the researchers conducted a series of experiments using diverse software systems and fault scenarios. The research compared the performance of the adaptive algorithm with traditional fault detection methods, including rule-based techniques and fixed neural network architectures. Evaluation metrics such as accuracy, precision, recall, and F1 score were used. The results consistently show that the adaptive neural transfer learning algorithm outperforms existing methods, achieving higher fault detection accuracy and improved classification performance.

Author 1: Jasem Alostad

Keywords: Software fault detection; adaptive neural algorithm; software reliability; neural networks; fault classification

Paper 10: Construction and Characteristics of an Engineering Economic Risk Management Platform Based on the BO-GBM Model

Abstract: Economic risk control is pivotal to the success of engineering projects. Traditional risk assessment methods often fall short in handling the high-dimensional, nonlinear, and strongly correlated risk factors prevalent in modern large-scale projects. To address these limitations, this study constructs an engineering economic risk management platform based on the BO-GBM model, which integrates Bayesian Optimization (BO) with a Gradient Boosting Machine (GBM). The platform employs a systematically constructed four-dimensional feature system encompassing 28 indicators across project ontology, market environment, execution process, and risk association dimensions. A rolling time window strategy is adopted for dynamic model training. Experimental validation on a dataset of 327 projects demonstrates the superior performance of the BO-GBM model: for classification tasks, it achieves an AUC of 0.927 and a recall rate of 91.3%, outperforming the standard GBM by 17.5 percentage points in recall; for regression tasks (cost deviation prediction), it attains an RMSE of 83,200 RMB and reduces the MAPE to 9.7%, surpassing mainstream baseline models. The platform's layered architecture (data, model, service, application layers) enables efficient risk identification and early warning: the time required for risk identification in large projects is drastically reduced from 42.6 hours to 0.52 hours, representing an 81.9-fold efficiency gain; the average single prediction response time is below 127 milliseconds, with a P95 response time of 427 milliseconds under 500 concurrent users; the early warning accuracy reaches 72.5%, with high-risk warnings issued up to 28 days in advance for cost risks and 42 days for schedule risks.

Author 1: Chaojian Wang

Author 2: Die Liu

Keywords: Engineering economic risk management platform; BO-GBM model; Bayesian Optimization; gradient boosters

Paper 11: Adaptive Virtual Machine Consolidation Based on Autoformer and Enhanced Double Q-Network for Energy-Efficient Cloud Data Center

Abstract: As the scale of cloud data centers continues to expand, energy consumption has become a critical issue. Virtual machine (VM) consolidation is a key technology for improving resource utilization and reducing energy consumption, yet it remains challenging to effectively balance energy efficiency with service level agreement violations (SLAV) in dynamic cloud environments. This paper proposes an adaptive VM consolidation strategy based on Autoformer and an enhanced dual Q-Network, referred to as AEDQN-VMC. The approach consists of three integrated components: 1) Autoformer-based load detection, which leverages an autocorrelation mechanism to decompose time-series data into multi-scale trend and periodic components; 2) a VM selection method that integrates the Pearson correlation coefficient and migration time to optimize the selection of VMs for migration; and 3) an enhanced dual Q-Network for VM placement, incorporating the upper confidence bound (UCB) and adaptive learning rate (ALR) to improve the exploration-exploitation trade-off. Extensive experiments on real-world cloud workload traces (PlanetLab, Google Cluster, and Alibaba datasets) demonstrate that the proposed method significantly outperforms state-of-the-art benchmarks such as PABFD, AD-VMC, and AMO-VMC. Specifically, it achieves maximum reductions of 46.5% in energy consumption and 74.2% in SLAV rate. Ablation studies further validate the contribution of each component and confirm the synergistic effect of the overall architecture. The results highlight the potential of AEDQN-VMC as an efficient and reliable solution for sustainable cloud data center operations.

Author 1: Kaiqi Zhang

Author 2: Youbo Lyu

Author 3: Dequan Zheng

Author 4: Yanping Chen

Author 5: Jianshan Xu

Keywords: Cloud computing; virtual machine consolidation; load prediction; energy efficiency; deep reinforcement learning; Autoformer

Paper 12: Handwriting Detectives Using Wavelet Siamese Technology to Verify Signature Fraud

Abstract: This paper addresses the escalating challenge of signature forgery detection through an innovative hybrid verification system. We integrate Siamese Neural Networks with wavelet scattering transformations to precisely capture signature characteristics while accommodating inherent variations. Our principal contribution, the "common anchor methodology," identifies a singular representative signature per individual, substantially reducing computational demands on the CEDAR Dataset while maintaining verification integrity. Through meticulous optimization of wavelet scattering parameters, our system demonstrates markedly superior performance on the CEDAR benchmark while requiring considerably fewer model parameters than traditional CNN architectures. This research establishes noteworthy advancements in both accuracy and efficiency for practical signature verification implementations. The study evaluates the performance of a wavelet-Siamese network architecture for offline signature verification through a series of five experiments with varying parameter configurations. Key variables include the use of a common anchor, the J Factor, and the θ value. Results reveal that incorporating a common anchor consistently improves performance. Among all configurations, experiment 4 with a J Factor of 2 and a θ value of 16 yielded the most favorable results, achieving the lowest error rate of 20.823% and the highest ROC-AUC score of 0.8699, along with efficient convergence within 55 iterations. In contrast, the absence of a common anchor in Experiment 1 led to a notably higher error rate of 24.44% and lower model performance. These findings demonstrate the critical role of parameter tuning in enhancing the robustness and accuracy of signature verification systems based on Siamese networks. Despite the substantial computational savings, the system’s best achieved error rate (20.82%) remains higher than several state-of-the-art and commercial signature verification solutions, many of which report error rates below 10%. This indicates an existing trade-off between efficiency and the highest attainable accuracy, which future work will aim to mitigate.

Author 1: Mohamed Nazir

Author 2: Ali Maher

Author 3: Mostafa Eltokhy

Author 4: Ali M. El-Rifaie

Author 5: Tarek Hosny

Author 6: Hani M. K. Mahdi

Keywords: Biometric authentication; Siamese neural networks; scattering wavelets; common anchor selection; neutrosophic logic; signature verification

Paper 13: A New Hybrid Algorithm for Vision-Based Sleep Posture Analysis Integrating CNN, LSTM and MediaPipe

Abstract: Sleep posture is a critical factor affecting sleep quality and long-term health, particularly for the elderly and patients with chronic conditions. This research proposes a novel hybrid algorithm for real-time, vision-based sleep posture analysis by integrating Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and MediaPipe pose estimation. The primary objective is to accurately classify the four main sleep postures—supine, left lateral, right lateral, and prone—while incorporating an automated alert system for risky behaviors, such as maintaining a prone position for over 15 minutes or remaining in any static posture for more than 2 hours. The system processes video input through a streamlined pipeline: MediaPipe first extracts 3D body keypoints, which are then fed into a CNN for spatial feature extraction, followed by an LSTM network to model temporal dependencies across frames. Evaluated on a dataset of 280 video samples from 20 participants under both daytime and nighttime conditions, the model achieved an accuracy of 96.4% in daylight and 92.8% in low-light environments, demonstrating robust performance across varying illumination. Comparative analysis confirmed its superiority over existing methods, such as depth-based CNN or pressure-sensor models. The study concludes that the proposed hybrid system offers a practical, non-invasive, and highly accurate solution for continuous sleep monitoring, with significant potential for deployment in smart healthcare and remote elderly care applications.

Author 1: Apichaya Nimkoompai

Author 2: Puwadol Sirikongtham

Keywords: Sleep posture detection; MediaPipe; CNN; LSTM; real-time monitoring; pose estimation

Paper 14: NetDAIL: An Optimized Deep Learning-Based Hybrid Model for Anomaly Detection in Network Traffic

Abstract: Detecting rare and subtle anomalies is critical for ensuring cybersecurity, financial integrity, and operational safety. High-dimensional features, severe class imbalance, and large data volumes often challenge conventional intrusion detection methods. This study presents NetDAIL, a hybrid framework that integrates deep feature learning using a denoising autoencoder, anomaly scoring through Isolation Forest, and classification via LightGBM to address these challenges. To evaluate its effectiveness, the proposed framework was tested on two widely used benchmark datasets: NSL-KDD for controlled-scale experimentation and KDD Cup 1999 for large-scale evaluation. NetDAIL achieved an AUC of 0.998 on the NSL-KDD dataset and 0.990 on the KDD Cup 1999 dataset, demonstrating strong discriminative capability across different traffic volumes and attack patterns. Experimental results confirm the model’s high detection accuracy, scalability, and generalization across diverse network intrusion scenarios. These findings highlight NetDAIL as a practical and reliable solution for real-world anomaly detection, capable of efficiently handling both small- and large-scale environments while maintaining robust and effective performance in operational settings.

Author 1: Saad Khalifa

Author 2: Mohamed Marie

Author 3: Wael Mohamed

Keywords: Anomaly detection; deep learning; autoencoders; NetDAIL; unsupervised learning; intrusion detection; NSL-KDD; KDD Cup 1999

Paper 15: An Integrated CNN, YOLOv5 and Faster R-CNN Framework for Real-Time Water Pipe Defect Detection

Abstract: In the context of rapidly expanding urban water supply networks and the prevalence of pipe defects – for example, corrosion, cracks, leaks, blockages – that undermine efficiency and pose safety risks, this study presents an intelligent detection system aimed at improving maintenance accuracy and operational stability. We propose a fusion-based detection architecture combining Convolutional Neural Networks for stable multi‐level feature extraction, YOLOv5 for high‐speed real‐time detection, and Faster R‐CNN for enhanced recall of small or occluded defects. Individually, the models achieve 85.0% accuracy for the CNN extractor, 90.0% detection accuracy with 50 FPS for YOLOv5, and 86.8% recall for Faster R‐CNN. Ablation experiments confirm that the fully integrated system attains superior performance—92.1% accuracy, 85.0% recall, an F1 score of 81.0, and an mAP of 85.1 at 45 FPS—demonstrating that ensemble methods harness complementary strengths to optimize detection precision and speed. Overall, our findings highlight the promise of deep learning–based ensembles for large‐scale, real‐time pipeline inspection, offering a foundation for future intelligent infrastructure management.

Author 1: Chu Fu

Author 2: Mideth Abisado

Keywords: Deep learning; Convolutional Neural Network; YOLOv5; Faster R-CNN; machine vision

Paper 16: SD-CNN: A Novel Lightweight Convolutional Neural Network Model for Fall Detection

Abstract: Aiming at the traditional deep learning fall detection model due to high computational complexity and a large number of parameters, this study proposes a lightweight convolutional neural network model, SD-CNN (SMA-Enhanced Depthwise Convolutional Neural Network), for fall detection. The model is first designed with an SMA attention module to enhance feature representation. Then, depth separable convolution is used to significantly reduce the model complexity. Finally, batch normalisation and Dropout regularisation techniques are combined to efficiently extract spatial-temporal features from temporal signals for accurate classification of fall and non-fall behaviours. The experiments use a sliding window to extract discrete features, three-axis acceleration, and synthetic acceleration as feature inputs. SD-CNN achieves 99.11% accuracy, 98.78% specificity, and 99.39% sensitivity on the homemade dataset Act, which are improved by 7.14%, 6.42%, and 9.38%, respectively, compared to CNN, while the number of parameters is reduced significantly. The effectiveness of the model is also verified by generalisation experiments on the public datasets SisFall and WEDAFall. The SD-CNN algorithm can efficiently complete the fall detection task, and the lightweight design makes it particularly suitable for wearable devices, which provides a highly efficient and reliable solution for real-time fall detection, and it has an important value for practical applications.

Author 1: Han-lin Shen

Author 2: Tian-hu Wang

Author 3: Hong Mu

Keywords: Fall detection; lightweight; SMA attention; depth-separable convolution

Paper 17: Efficient Lightweight Detection and Classification Method for Field-Grown Horticultural Crops

Abstract: As the core carrier of human food supply and agricultural economy, manual management in large-scale crop cultivation faces bottlenecks such as low efficiency, high cost, and difficulty in standardization. There is an urgent need for computer vision technology to realize automated detection and growth stage classification. However, most existing algorithms rely on high-performance GPUs for operation, resulting in high hardware costs, which makes it difficult to popularize them in low-end agricultural edge devices (e.g., embedded controllers, low-cost industrial computers). This study proposes a lightweight crop detection and classification model, Lite-CropNet. It builds a neural network architecture based on the CSPDarknet backbone network, designs a concise decoder, and adopts four-scale detection heads to adapt to crop targets of different sizes, balancing high accuracy and lightweight characteristics. Using tomatoes as the experimental object, tests on the TomatOD dataset (simulating real greenhouse environments) show that Lite-CropNet outperforms advanced methods, with a mean Average Precision (mAP)@0.5 of 85.7%. Under the conditions of the GTX 1650 GPU and 640×640 resolution, the Frame Per Second (FPS) reaches 76.9, and the model size is only 4.4M. This neural network model can efficiently complete tomato detection and maturity classification, and its architecture and design can also be transferred to crops such as potatoes and strawberries, providing a cost-effective and highly universal automated solution for agricultural production.

Author 1: Yaru Huang

Author 2: Hua Zhou

Author 3: Zhongyi Shu

Keywords: Computer vision; neural network; object detection and classification; lightweight; horticultural crops

Paper 18: Using Combined Weighting and BP Neural Networks for Relative Poverty Measurement and its Evaluation

Abstract: This study addresses the challenges of measuring and evaluating relative poverty by introducing a comprehensive evaluation model based on the Analytic Hierarchy Process (AHP)-entropy method and BP neural networks. A multidimensional evaluation index system was constructed through expert consultation and literature review. The AHP-entropy method was then employed to determine the weights of the evaluation indicators, ensuring objectivity and scientific validity. Additionally, the BP neural network model was integrated to leverage self-learning and adaptive mechanisms for efficient and accurate poverty assessment. Empirical analysis shows that the model maintains a calculation error within 3.9%, demonstrating high precision and wide applicability. This research provides a novel approach that combines qualitative analysis with quantitative evaluation, offering a practical tool for governmental agencies to design effective poverty alleviation strategies. Moreover, the model opens new pathways for future research in regional poverty assessment, especially in enhancing cross-cultural adaptability and advancing intelligent evaluation models.

Author 1: Xiaohua Cai

Author 2: Ya Zhao

Author 3: Lijia Chen

Author 4: Juan Huang

Author 5: Yang Xu

Keywords: Analytic hierarchy process (AHP); entropy method; BP neural network model; relative poverty measurement

Paper 19: Evaluating Transformer-Based Pretrained Models for Classical Arabic Named Entity Recognition

Abstract: This study presents a comprehensive comparative evaluation of transformer-based pretrained language models for Named Entity Recognition (NER) in Classical Arabic, an underexplored linguistic variety characterized by rich morphology, orthographic ambiguity, and the absence of diacritics. The main objective of this work is to identify the most effective transformer model for Classical Arabic NER and to analyze the linguistic factors influencing model performance. Using the CANERCorpus, which contains Hadith texts annotated with twenty fine-grained entity types, ten transformer-based models were fine-tuned and evaluated under consistent experimental settings. The study benchmarks models such as AraBERT, ArBERT, and multiple CAMeLBERT variants, comparing their precision, recall, and F1-scores. The results demonstrate that all models achieve strong performance (F1 > 96%), while CAMeL-CA-NER attains the highest score (F1 = 97.78%), confirming the advantage of domain-specific pretraining on Classical Arabic data. Error analysis further reveals that domain-adapted models better handle ambiguous entities and religious terminology. A comparative analysis with traditional and non-transformer approaches, including rule-based and BERT-CRF models from previous studies, shows that CAMeL-CA-NER surpasses earlier methods by more than 3% in F1-score, highlighting its superior capability in handling Classical Arabic text. However, this study is limited to the CANERCorpus, which primarily consists of Hadith texts; results may vary for other Classical Arabic genres or domains. These findings provide a valuable benchmark for future research and demonstrate the adaptability of modern NLP architectures to linguistically complex, low-resource domains.

Author 1: Mariam Muhammed

Author 2: Shahira Azab

Keywords: Classical Arabic; Named Entity Recognition; transformer models; pretrained models; CANERCorpus

Paper 20: A Hippopotamus Optimization Algorithm-Based Convolutional Neural Network Model for Mental Health Assessment Among College Students

Abstract: The mental health of adult students is crucial not only for enhancing their learning experience and overall quality of life, but also for alleviating academic and employment-related anxiety. A significant challenge in developing effective online mental health support systems is the accurate assessment of students' mental health status. Current evaluation methods often lack precision and fail to integrate multifaceted data perspectives. To address these challenges, this study developed a psychological assessment system based on deep learning technology. The system aims to assess adult students' psychological states and provide appropriate support. Specifically, it utilizes a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) algorithm framework to evaluate students' psychological states by synthesizing image data, academic performance, and textual inputs. Furthermore, to enhance the accuracy of deep learning-based mental health assessment models, an improved hippopotamus optimization (IHO) algorithm was designed to optimize the hyperparameters of deep learning frameworks. By using the proposed multi-input single-output hybrid IHO-based LSTM-CNN framework (IHO-LSTM-CNN), the online mental health assessment module can accurately describe the psychological status of college students and provide personalized support to meet their specific needs. The final results indicate that the IHO-LSTM-CNN framework provides more accurate assessments than existing mental health assessment models, with an accuracy of 90.28%. This enhanced accuracy enables online community psychological support systems to deliver precise and effective psychological support to college students.

Author 1: Gai Hang

Author 2: Lin Yang

Keywords: Convolutional Neural Network; Long Short-Term Memory; hippopotamus optimization algorithm; mental health assessment; deep learning

Paper 21: A Hybrid Deep Learning and IoT Framework for Predictive Maintenance of Wind Turbines: Enhancing Reliability and Reducing Downtime

Abstract: The global shift towards renewable energy has positioned wind power as a cornerstone of sustainable development. However, the operational efficiency of wind farms is significantly hampered by unexpected component failures, leading to substantial downtime and maintenance costs. Traditional scheduled maintenance protocols are inefficient, often leading to unnecessary interventions or catastrophic failures. This study proposes a novel, robust framework for the predictive maintenance (PdM) of wind turbines, integrating Internet of Things (IoT) sensory data with a hybrid deep learning architecture. The proposed model leverages Convolutional Neural Networks (CNN) for feature extraction from vibrational and acoustic emission data, combined with Long Short-Term Memory (LSTM) networks to model the temporal dependencies inherent in time-series operational data. Drawing inspiration from successful applications of similar hybrid AI models in precision agriculture and smart farming, our approach is designed to accurately forecast the Remaining Useful Life (RUL) of critical components like gearboxes and bearings. We validate our framework on a benchmark dataset from NASA's Pronostia platform, demonstrating a 30% improvement in prediction accuracy over traditional single-model approaches and a 50% reduction in false alarms. The results underscore the potential of integrating hybrid AI and IoT, a paradigm successfully demonstrated in other complex systems, to create more reliable, efficient, and cost-effective maintenance strategies for the wind energy sector, thereby enhancing grid stability and accelerating the renewable energy transition.

Author 1: Amina Eljyidi

Author 2: Hakim Jebari

Author 3: Siham Rekiek

Author 4: Kamal Reklaoui

Keywords: Predictive maintenance; wind turbine; artificial intelligence; deep learning; Convolutional Neural Network; Long Short-Term Memory; Internet of Things; Remaining Useful Life; condition monitoring

Paper 22: Federated Machine Learning for Monitoring Student Mental Health in Kazakhstan

Abstract: Federated Learning (FL) offers a privacy-preserving and decentralized paradigm for machine learning, making it particularly suitable for analyzing sensitive psychological and physiological data. This study aims to develop and evaluate a federated learning framework for assessing the psycho-emotional well-being of students in Kazakhstani educational institutions, where data privacy and infrastructural constraints pose significant challenges. We benchmark three FL algorithms, such as FedAvg, FedOpt, and FedProx, on heterogeneous, institution-level datasets that combine sleep, dietary, activity, and self-reported emotional measures. Experiments simulate cross-device, non-IID deployments and evaluate convergence, accuracy, and stability across ten communication rounds. Results show that FedProx attains the best trade-off between accuracy and stability under non-IID conditions (peak accuracy is 99.9%), while FedOpt provides faster early convergence, and FedAvg performs well for more homogeneous partitions. The methodological contribution comprises optimized aggregation and adaptive client weighting to mitigate non-IID effects in resource-constrained educational settings. These findings validate FL as a scalable, privacy-preserving approach for mental health monitoring in education and support its use for early intervention and resilience tracking. The proposed framework contributes to data-driven mental health policy design in educational systems, addressing both ethical and infrastructural considerations. The study discusses limitations of the simulated setup and outlines directions for broader deployment and cross-silo validation.

Author 1: Bakirova Gulnaz

Author 2: Bektemyssova Gulnara

Author 3: Nor'ashikin Binti Ali

Keywords: Federated Learning; data privacy; FedOpt; FedAvg; FedProx; mental health; non-IID data; educational data mining; psychological analytics

Paper 23: Exploring Hallucination in Large Language Models

Abstract: Large Language Models such as GPT-4o and GPT-4o-mini have shown significant promise in various fields. However, hallucination, when models generate inaccurate information, remains a critical challenge, especially in domains that require high accuracy, such as the healthcare field. This study investigates hallucinations in two different LLMs, focusing on the healthcare domain. Four different experiments were defined to examine the two models’ memorization and reasoning abilities. For each experiment, a dataset with 193,155 multiple-choice medical questions from postgraduate medical programs was prepared by splitting it into 21 subsets according to medical topics. Each subset has two versions: one with the correct answers included and one without them. Accuracy and compliance were evaluated for each model. Models’ adherence to requirements in prompts was assessed. Also, the correlation between size and accuracy was tested. The experiments were repeated to evaluate the models’ stability. Finally, the models’ reasoning was evaluated by human experts who assessed the models’ explanations for correct answers. The results revealed poor rates of accuracy and compliance for the two models, with rates below 70% and 75%, respectively, in most datasets; yet, both models showed low uncertainty (3%) in their responses. The findings showed that the accuracy was not affected by the size of the dataset provided to the models. Also, the results indicated that GPT-4o-mini demonstrates greater performance stability compared to GPT-4o. Furthermore, the two models provided acceptable justifications for choosing the correct answer in most cases, according to 68.8% of expert questionnaire participants who agreed with both models’ justifications. According to these results, both models cannot be relied upon when accuracy is critical, even though GPT-4o-mini slightly outperformed GPT-4o in providing the correct answers. The findings highlight the importance of improving LLM accuracy and reasoning to ensure reliability in critical fields like healthcare.

Author 1: Nesreen M. Alharbi

Author 2: Thoria Alghamdi

Author 3: Raghda M. Alqurashi

Author 4: Reem Alwashmi

Author 5: Amal Babour

Author 6: Entisar Alkayal

Keywords: ChatGPT; GPT-4o; GPT-4o-mini; hallucination; healthcare; large language models

Paper 24: Game-Theoretic Approaches for Robust Stability of DC Motor Systems

Abstract: This study proposes a game-theoretic framework for achieving robust stability in DC motor systems operating under parametric uncertainty and external disturbances. We model the controller, disturbance, and uncertainty as strategic players in a non-cooperative differential game and synthesize equilibrium policies using a Lyapunov–game approach. Practically, the method integrates: 1) LMI-based stabilization to certify descent conditions, 2) actor–critic reinforcement learning to approximate the Hamilton–Jacobi–Isaacs (HJI) value function beyond linear regimes, and 3) evolutionary/swarm optimization for controller initialization and distributed observer tuning. We validate the framework on a separately excited DC motor subject to ±20% parameter variations and a bounded load-torque disturbance and compare it against PID and H8 baselines. Simulations show consistently faster rise/settling, lower overshoot, stronger disturbance rejection at a step disturbance, and smoother control effort, while attaining the highest qualitative robustness margin among the tested controllers. Beyond single-motor stabilization, we outline extensions to multi-agent coordination, security-aware control, and fractional/fuzzy models, demonstrating adaptability and scalability of the approach. These results indicate that framing stability as the outcome of strategic interactions yields reliable and efficient DC-motor control in uncertain, adversarial environments.

Author 1: Mohamed Ayari

Author 2: Atef Gharbi

Author 3: Yamen El Touati

Author 4: Zeineb Klai

Author 5: Mahmoud Salaheldin Elsayed

Author 6: Elsaid Md. Abdelrahim

Keywords: Game theory; DC motor control; robust stability; differential games; Lyapunov stability; reinforcement learning; evolutionary algorithms

Paper 25: Comparative Performance Analysis of Original AuRa and Improved AuRa Consensus Algorithms in Chain Hammer Digital Certificate Simulation

Abstract: The blockchain functions as a distributed database, where data is securely stored across multiple servers and network nodes. It exists in various forms, with Bitcoin, Ethereum, and Hyperledger being among the most prominent examples. To ensure the integrity and security of transactions within a blockchain network, a consensus algorithm is employed to establish agreement among participating nodes. Several types of consensus algorithms exist, each offering distinct features and operational mechanisms. One such algorithm is Authority Round (here defined as AuRa_ori), a member of the Proof-of-Authority (PoA) family supported by Parity clients. Previous studies have highlighted several vulnerabilities and performance limitations in AuRa_ori, particularly concerning transaction speed per second (TPS) and transaction throughput per second (TGS). This study specifically investigates the original AuRa algorithm alongside an improved version, termed AuRa_v1. In AuRa_v1, the transaction process is structured into four key phases: 1) leader assignment, 2) block proposal, 3) agreement, and 4) block commitment. However, inconsistencies and inefficiencies have been identified within certain phases of the original AuRa_ori, particularly during the leader assignment and agreement stages. In response, this study proposes an improved approach through AuRa_v1 to address these vulnerabilities. A detailed analysis is conducted to evaluate the impact of these vulnerabilities on TPS, TGS, and epoch time, followed by a performance comparison between AuRa_ori and AuRa_v1. Experimental results demonstrate that AuRa_v1 effectively resolves the identified performance issues, achieving a significant improvement. Specifically, AuRa_v1 records a 21.65% increase in both TPS and TGS compared to AuRa_ori, validating the effectiveness of the proposed enhancements.

Author 1: Robiah Arifin

Author 2: Wan Aezwani Wan Abu Bakar

Author 3: Mustafa Man

Author 4: Evizal Abdul Kadir

Keywords: Blockchain; Ethereum; AuRa_ori; AuRa_v1; TPS; TGS

Paper 26: An Improved Marine Predators Algorithm-Based UAV Path Planning for 10-kV Distribution Networks Inspection in Live Working Scenarios

Abstract: Before conducting maintenance on 10-kV distribution networks, the use of unmanned aerial vehicles (UAVs) for inspecting distribution lines can effectively enhance the operational efficiency of personnel in live working scenarios. For UAV-based inspection of power distribution networks, an optimal flight path ensures both operational safety and comprehensive image acquisition in live working scenarios. Therefore, this study proposes a UAV path planning algorithm and an insulator defect classification model based on YOLOv11, aiming to develop a UAV system for live power line detection. Firstly, a UAV path planning model is established to minimize the flight path length and maximize the image acquisition range, which also considers the safety distance constraints between UAVs and live power lines. On this basis, the optimization strategy of the particle swarm optimization (PSO) algorithm is introduced into the marine predictors algorithm (MPA), and a hybrid PSO-MPA algorithm is designed to improve the convergence accuracy of the MPA algorithm and solve the proposed UAV planning model. In addition, an insulator defect detection model has been developed to accurately identify the image information collected by UAVs. In order to improve the accuracy of the YOLOv11 model, the task-separation assignment (TSA) module was introduced into the YOLOv11 model, and a TSA-YOLOv11 model was designed. Experimental results demonstrate that the proposed PSO-MPA algorithm achieves superior convergence accuracy compared to five algorithms, including PSO. When the UAV flight step size is one meter, the PSO-MPA algorithm reduces the objective function value by an average of 49.62% relative to the other algorithms. Additionally, the TSA-YOLOv11 model attained an average accuracy of 96.87% for the insulator defect classification problem.

Author 1: Dapeng Ma

Author 2: Hongtao Jiang

Author 3: Lichao Jiang

Author 4: Chi Zhang

Author 5: Changwu Li

Author 6: Xin Zheng

Author 7: Mingxian Liu

Author 8: Kai Li

Keywords: Marine predictors algorithm; YOLOv11; defect classification; UAV path planning; live power lines

Paper 27: Enhancing Predictive Maintenance Method Using Machine Learning to Improve IoT-Embedded Machinery Efficiency and Performance

Abstract: Predictive maintenance plays a crucial role in minimizing unplanned downtimes, reducing maintenance costs, and optimizing the operational efficiency of IoT-embedded industrial machinery. Despite its transformative potential, traditional predictive maintenance methods often face challenges such as limited accuracy, high latency, and inefficiencies in processing large and imbalanced datasets. This study proposes an enhanced predictive maintenance method using the Sliding Window Method with XGB model (E.XGB), incorporating advanced data preprocessing, permutation importance, and hyperparameter optimization to address these limitations. The proposed method was evaluated on two datasets, which are the synthetic AI4I 2020 Predictive Maintenance Dataset and the real-world CNC Milling Dataset. A comparative analysis with a predictive maintenance method using E.AB from prior research as a benchmark, along with several baseline models, DT, RF, and SVM, revealed that the E.XGB model consistently outperformed other methods in accuracy, precision, recall, and F1-scores. On the AI4I2020 dataset, the E.XGB model achieved an accuracy of 99.05%, while on the CNC Milling dataset, it attained an accuracy of 99.01%. Additionally, the E.XGB model also demonstrated reduced training and prediction times, meeting the real-time requirements of industrial applications. The proposed model demonstrated training speed of approximately 94% and prediction speeds of approximately 99.8% improvement over the E.AB model, making it highly suitable for real-time industrial applications. By improving accuracy, training speed, and prediction latency, the predictive maintenance method offers a robust, scalable, and reliable solution for predictive maintenance across diverse industrial contexts.

Author 1: Abiinesh Nadarajan

Author 2: Iskandar Ishak

Author 3: Noridayu Manshor

Author 4: Raihani Mohamed

Author 5: Mohamad Yusnisyahmi Yusof

Keywords: Internet of Things; machine learning; predictive maintenance

Paper 28: A Hybrid Deep Learning and Forensic Approach for Robust Deepfake Detection

Abstract: The rapid evolution of generative adversarial networks (GANs) and diffusion models has made synthetic media increasingly realistic, raising societal concerns around misinformation, identity fraud, and digital trust. Existing deepfake detection methods either rely on deep learning, which suffers from poor generalization and vulnerability to distortions, or forensic analysis, which is interpretable but limited against new manipulation techniques. This study proposes a hybrid framework that fuses forensic features—including noise residuals, JPEG compression traces, and frequency-domain descriptors—with deep learning representations from convolutional neural networks (CNNs) and vision transformers (ViTs). Evaluated on benchmark datasets (FaceForensics++, Celeb-DF v2, DFDC), the proposed model consistently outperformed single-method baselines and demonstrated superior performance compared to existing state-of-the-art hybrid approaches, achieving F1-scores of 0.96, 0.82, and 0.77, respectively. Robustness tests demonstrated stable performance under compression (F1 = 0.87 at QF = 50), adversarial perturbations (AUC = 0.84), and unseen manipulations (F1 = 0.79). Importantly, explainability analysis showed that Grad-CAM and forensic heatmaps overlapped with ground-truth manipulated regions in 82 per cent of cases, enhancing transparency and user trust. These findings confirm that hybrid approaches provide a balanced solution—combining the adaptability of deep models with the interpretability of forensic cues—to develop resilient and trustworthy deepfake detection systems.

Author 1: Sales Aribe Jr

Keywords: Adversarial robustness; deepfake detection; diffusion models; explainable AI; forensic fusion; multimedia forensics; trustworthy AI

Paper 29: Facial Expression Recognition Under Partial Occlusion Using Part-Based Ensemble Learning

Abstract: Facial expression recognition (FER) under partial occlusion remains a challenging task, especially when key regions of the face, such as the mouth and nose, are covered by medical masks. Such conditions significantly reduce the discriminative features available for accurate emotion recognition, limiting the effectiveness of conventional full-face approaches. To address this issue, this study proposes a part-based learning framework that partitions the face into multiple regions, allowing the model to exploit unoccluded areas for expression recognition. The proposed method employs Support Vector Machine (SVM) classifiers trained on Histogram of Oriented Gradients (HoG) features extracted from 2, 3, 4, and 6 facial partitions. Each part-based model is trained independently, and their outputs are combined through a weighted soft voting ensemble mechanism to generate the final prediction. The experiments were conducted on the MaskedFER2013 dataset, which contains 31,116 grayscale facial images (48×48 pixels) distributed across seven emotion classes. The results demonstrate that the four-part model achieves the best performance, reaching an accuracy of 45%, outperforming both single-part models and full-face baselines under occlusion scenarios. These findings confirm that the proposed part-based ensemble approach enhances the robustness of FER systems by effectively leveraging complementary regional features, thereby providing a promising solution for real-world applications, where facial occlusion is unavoidable.

Author 1: Evangelions Felix Yehdeya

Author 2: Wahyono

Keywords: Facial expression recognition; partial occlusion; partial part model; support vector machine; ensemble learning

Paper 30: Ambulance Detection and Priority Passage at Urban Intersections Using Transfer Learning and Explainable AI

Abstract: Static traffic signal timings often cause severe delays for emergency vehicles, including ambulances at junctions in urban areas, putting lives at risk. To highlight this, the present study proposes an intelligent traffic control system that dynamically adjusts traffic signals based on real-time monitoring. The system employs a yolov8-based deep learning model fine-tuned through transfer learning for ambulance detection from live video. At an Intersection over Union (IoU) threshold of 0.5, the model achieves a mean Average Precision (mAP) of 0.860. To ensure continuous tracking, NORFair tracking is implemented to ensure constant detection across frames. Additionally, to improve explainability and, the frame incorporates Local Interpretable Model-Agnostic Explanation (LIME), providing visual signals into the model decision-making process. Once an ambulance is detected, the system instantly triggers a green-light activation for the ambulance's lane, enabling quick emergency response. Unlike conventional systems with fixed signal timing, this approach enables smart and adaptive traffic management in urban environment. However, the system's shortcomings in low-visibility situations, such as at night or in fog, despite its encouraging results, highlight the need for incorporating images taken at night and in foggy weather into the dataset.

Author 1: Murtaza Hanif

Author 2: Taj Muhammad

Author 3: Atif Ikram

Author 4: Shahid Yousaf

Author 5: Marwan Abu-Zanona

Author 6: Asef Mohammad Ali Al Khateeb

Author 7: Bassam Elzaghmouri

Author 8: Saad Mamoun Abdel Rahman Ahmed

Author 9: Lamia Hassan Rahamatalla

Keywords: Ambulance detection; YOLOV8; LIME; transfer learning; NorFair; urban area; traffic control; smart traffic management

Paper 31: Digital Trust and Legacy: Mapping the Intersection of Inheritance Systems and Emerging Technologies (2010–2025)

Abstract: Inheritance systems worldwide are undergoing a paradigm shift evolving from manually administered processes to technologically enabled platforms for managing both tangible and digital assets. Yet, the scholarly understanding of how technologies ranging from information systems to blockchain have transformed inheritance management remains underexplored and fragmented. This study aims to trace the evolution of inheritance systems from 2010 to 2025, with a particular focus on the digitalization of inheritance management, emerging technologies and governance models. Using a bibliometric approach, 229 documents were initially retrieved from the Scopus database. After removing irrelevant records, a refined dataset of 81 publications was analyzed using Excel and VOSviewer. The analysis included performance metrics (e.g., publication growth, citation trends, and country output) and science mapping (keyword co-occurrence and clustering). Findings reveal a significant rise in publications post-2020, coinciding with increased attention to digital assets, data privacy laws (e.g., GDPR) and emerging technologies such as blockchain. The most active contributors were from the United States, China and the United Kingdom. Highly cited articles discuss themes such as digital legacy, legal frameworks, asset authentication and ethical considerations. Thematic clustering revealed four research domains: digital legacy and estate transition, digital transformation and trust, digital asset structuring and fraud prevention in social media inheritance. This study contributes a comprehensive overview of the field’s conceptual landscape by highlighting the uneven yet accelerating integration of digital tools in inheritance systems. It also underscores the urgent need for inclusive, interdisciplinary frameworks that accommodate diverse legal, cultural and technological contexts for future inheritance governance.

Author 1: Nor Aimuni Md Rashid

Author 2: Faiqah Hafidzah Halim

Author 3: Hazrati Zaini

Author 4: Norshahidatul Hasana Ishak

Author 5: Nur Farahin Mohd Johari

Author 6: Alya Geogiana Buja

Keywords: Inheritance systems; digitalization; secure data; trust; technologies; digital legacy; blockchain; digital assets

Paper 32: A Quality Assessment Study of Deep Learning Techniques for Medical Image Diagnosis and Their Applications: A Systematic Literature Review

Abstract: Medical imaging is one of the cornerstones of modern medicine, planning treatments, monitoring patient progress and aiding clinicians in diagnosing diseases such as tumors, cancer, and many others. With the rise of neural networks, especially deep learning (DL) approaches, significant advancements have been made in this domain. This systematic literature review intended to investigate and identify the latest implementations of DL algorithms for medical image processing by examining 294 peer-reviewed articles. We also explored the DL-based image segmentation methods, highlighting their advantages and limitations and the commonly used datasets in the field. Finally, we analyzed key challenges and outlined future research directions related to image segmentation. Our review reveals that convolutional neural networks, particularly U-Net and its variants, dominate the field, while deep neural networks show promising results enabling end-to-end learning, providing greater flexibility, and facilitating transfer learning. This study is conducted by defining the search process designed for execution based on a set of inclusion and exclusion criteria from major databases including IEEE explore, Scopus and DBLP.

Author 1: Amine Berquedich

Author 2: Ahmed Zellou

Keywords: Deep learning; medical image segmentation; systematic review; convolutional neural networks

Paper 33: Modeling and Analyzing Malware Behavior in Virtual Networks Using EVE-NG

Abstract: Malicious attacks have become increasingly common in all organizations and systems. The continued evolution of such software aims to extract information from diverse systems. Therefore, the objective of this study is to introduce another approach to analyze some network attacks, within a virtual infrastructure, through multi-vendor network emulation software (Emulated Virtual Environment–Next Generation - EVE-NG). Basically, through emulated resources, the aim is to implement a complex network, which also includes a Security Information and Event Management System (SIEM), which can detect some attacks, both from the network area (carried out by malicious attackers) and through malicious files (from the public resources area), that are accidentally or intentionally downloaded by certain users. Within this environment, various scenarios can be implemented to simulate the real production environment, in order to test network vulnerabilities, but also to improve some methods for learning network attack and defense modes. In the experiments performed, the SIEM system detected most of the simulated attacks, but failed to distinguish between the displayed alarms so that the alerts could indicate the type of attack. Thus, the potential of EVE-NG for simulating and analyzing the behavior of malware is demonstrated.

Author 1: Maria-Madalina Andronache

Author 2: Alexandru Vulpe

Author 3: Corneliu Burileanu

Keywords: EVE-NG; network attacks; network defense; SIEM; network architecture

Paper 34: Predicting Stock Market Performance Based on Sentiment Analysis of Online Comments

Abstract: In China's retail-focused stock market, the influence of social media sentiment during off-hours on the next day's opening price has received limited attention. This paper takes Kweichow Moutai—a leading Chinese company with substantial market capitalization—as the research sample. It gathers investor commentary data from financial platforms, and uses natural language processing tools (SnowNLP) to develop a multidimensional sentiment index (including average sentiment score, positive ratio, and sentiment volatility). By integrating this index with stock trading data and macroeconomic indicators, this study designs a dual-channel LSTM model: one channel for market technical features (e.g., price, volume) and the other for sentiment features, aiming to analyze the impact of off-hours sentiment on opening prices. Empirical results indicate that overnight sentiment has significant predictive power for the next day's opening price; meanwhile, sentiment transmission is asymmetric, making predictions more challenging in declining markets. Additionally, high-frequency sentiment data significantly outperforms low-frequency data in market prediction accuracy. This research expands the understanding of how investor sentiment influences the market over time, providing practical insights for market participants to develop effective strategies and manage risks.

Author 1: Wenhao Suo

Author 2: Tongjai Yampaka

Keywords: Investor sentiment; non-trading hour sentiment; social media comments; dual-channel LSTM

Paper 35: User Satisfaction in AI-Driven Islamic Fintech: An Extended Technology Acceptance Model with Task–Technology Fit and Sharia Compliance

Abstract: The rapid development of digital financial services has transformed financial intermediation through improved access, transparency, and efficiency. In the Indonesian context, Islamic financial technology (fintech) offers an alternative aligned with Sharia principles, particularly through e-ijarah contracts that provide MSMEs with productive asset access without interest-bearing debt. This study aims to empirically evaluate the determinants of user satisfaction in adopting AI-based e-ijarah applications by extending the Technology Acceptance Model (TAM) with Task–Technology Fit (TTF), Sharia Compliance (SC), Trust in AI, and Perceived Risk (PR). A survey of 75 food and beverage MSMEs in East Java was analyzed using Partial Least Squares Structural Equation Modeling (PLS-SEM). The findings indicate that Perceived Ease of Use (PEOU) strongly influences Perceived Usefulness (PU), which in turn significantly affects Behavioral Intention (BI), Actual Use (AU), and User Satisfaction (EUCS). Trust in AI and TTF also play significant roles in enhancing PU and BI. Interestingly, SC shows a significant but negative effect on PU, highlighting a contextual gap between digital automation and perceptions of religious compliance. PR negatively impacts both BI and AU, while Age does not moderate usage behavior. The study contributes conceptually by integrating TAM, TTF, and Sharia compliance in a single framework, and practically by offering insights for fintech developers and regulators to improve system usability, trust, and compliance clarity.

Author 1: Mardiana Andarwati

Author 2: Sari Yuniarti

Author 3: Andriyan Rizki Jatmiko

Author 4: Firnanda Al-Islama Achyunda Putra

Author 5: Galandaru Swalaganata

Author 6: Ahmad Taufiq Andriono

Keywords: Task–Technology fit; sharia compliance; technology acceptance model; user satisfaction; AI; Islamic fintech; MSMEs

Paper 36: Multimodal Deep Learning for Tuberculosis Detection Using Cough Audio and Clinical Data with Health Acoustic Representations (HeAR)

Abstract: Tuberculosis (TB) remains a significant global health challenge, necessitating rapid and accessible screening methods. This study proposes a multimodal deep learning model for non-invasive TB detection by fusing acoustic features from cough sounds with clinical metadata. We utilize the pre-trained Health Acoustic Representations (HeAR) model as a powerful backbone to extract features from mel-spectrograms of cough audio. These acoustic features are combined with clinical data, including sex, age, and key symptoms through a late-fusion architecture. The model was trained and evaluated on a balanced dataset of 16,000 samples derived from the CODA TB DREAM Challenge dataset. Our proposed multimodal approach achieved a high overall accuracy of 90% on the unseen test set, with balanced precision, recall, specificity, and F1-scores of 0.90 for both TB-positive and non-TB classes. These results demonstrate the effectiveness of using cough sound as a non-invasive vocal biomarker, amplified by combining advanced acoustic representations with clinical context. This highlights the potential of our method as a robust, low-cost, and scalable tool for early TB screening.

Author 1: Rinaldi Anwar Buyung

Author 2: Widi Nugroho

Keywords: Tuberculosis; cough detection; Health Acoustic Representation; multimodal; vocal biomarker

Paper 37: EYE-GDM: Clinically Validated, Explainable Ensemble Learning for Gestational Diabetes

Abstract: As artificial intelligence (AI) advances in healthcare, its use in maternal health shows promise but faces challenges of trust due to the black-box nature of many models. Gestational diabetes mellitus (GDM), a transient yet high-risk condition, demands accurate and interpretable prediction tools. However, existing GDM prediction studies often rely on opaque models or post-hoc explanation techniques applied after training, which limits transparency and reduces their clinical applicability. This highlights an urgent need for models that unify high predictive performance with interpretability by design. This study introduces EYE-GDM, a case-specific application of our Enhanced Interpretability Ensemble (EYE) framework, designed to predict GDM risk with clinically meaningful explanations. The pipeline evaluates multiple algorithms and selects Decision Tree (DT), k-Nearest Neighbors (k-NN), and Gradient Boosting (GB) as the best-performing base learners. These are integrated with SHAP and a logistic regression (LR) meta-model to construct EYE-GDM, embedding interpretability by weighting learner outputs with LR coefficients. This yields global (population-level) and local (patient-level) explanations consistent with medical knowledge. Tested on a dataset of 3,525 pregnancies, EYE-GDM achieved strong performance (accuracy = 0.9789, AUC-ROC = 0.9981) and provided insights into risk patterns, thresholds, and feature interactions relevant to GDM. By embedding explainability within the ensemble construction, EYE-GDM achieves transparent and clinically aligned reasoning without compromising predictive performance. Thus, EYE-GDM demonstrates how explainable AI (XAI) can translate from technical innovation to practical value in maternal care, supporting earlier risk identification and more informed clinical decisions.

Author 1: Shatha Alghamdi

Author 2: Rashid Mehmood

Author 3: Fahad Alqurashi

Author 4: Turki Alghamdi

Author 5: Sarah Ghazali

Author 6: Asmaa AlAhmadi

Keywords: Explainable Artificial Intelligence (XAI); interpretable machine learning (IML); Gestational diabetes mellitus (GDM); maternal health; healthcare AI; GDM risk prediction; transparency; trust

Paper 38: An Intelligent Platform for Behavior Modification and Office Syndrome Risk Reduction Using MediaPipe and Computer Vision

Abstract: Office Syndrome, a musculoskeletal disorder prevalent among office workers, poses significant risks to health, productivity, and quality of life. Traditional preventive approaches, such as ergonomic guidelines and reminder-based systems, often fail due to limited user adherence and practicality. To address this gap, this study developed an intelligent platform that integrates MediaPipe and computer vision to monitor sitting posture, eye-to-screen distance, and sitting duration in real time. The system provides automated notifications and stretching recommendations, combining detection, feedback, and behavioral intervention into a sensor-free and cost-effective solution. The platform was evaluated in terms of technical performance and user behavioral impact. Results demonstrated high system accuracy, with the eye-distance detection module achieving 95.2% accuracy, followed by long sitting alerts (92.5%) and proximity alerts (90.1%). User evaluations confirmed that real-time notifications increased awareness and encouraged healthier working behaviors. These findings highlight the potential of computer vision based approaches for ergonomic health promotion. The proposed platform contributes not only to preventive strategies for Office Syndrome but also to advancing user-centered, technology-driven health solutions adaptable to both office and remote work environments. This study not only demonstrates technical performance but also introduces a novel integration of Media Pipe based posture and facial detection with behavioral modification features, which previous ergonomic systems have not addressed. The proposed framework contributes a new perspective for integrating real time feedback with computer vision in promoting sustainable ergonomic behavior.

Author 1: Sumran Chaikhamwang

Author 2: Wijitra Montri

Author 3: Chalida Janthajirakowit

Author 4: Srinuan fongmanee

Keywords: Office syndrome; computer vision; MediaPipe; behavior modification; ergonomics

Paper 39: A Review of Ransomware Detection Models for Cybersecurity Driven IIoT in Cloud Environments

Abstract: Ransomware is currently one of the most severe cybersecurity threats and not only attacks legacy systems but cloud systems and Industrial Internet of Things (IIoT) systems as well. Security and privacy threats are heightened as these systems integrate more closely and thus are exposed to sophisticated and long-lasting attacks. This paper provides a comprehensive review of ransomware prevention and detection measures in cloud and IIoT environments with an emphasis on the usage of Machine Learning (ML) and Deep Learning (DL) models. Research studies published across IEEE, Elsevier, and Springer databases between 2020 and 2024 were analyzed. Our check reveals Ensemble methods and Random Forest (RF) are two of the ML methods most in use, with each at 18.00%, followed by Neural Networks (NNs) at 12.00%, with older models such as Support Vector Machines (SVMs) with 10.00%, Naïve Bayes (NBs) had 7.00%, and Decision Trees (DTs) still in use with utilization at 9.00% . Additionally, DL approaches (including Convolutional NN (NN), Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Recurrent NN (RNN)) account for 20.00% of the techniques deployed, highlighting their growing prominence in IIoT security and ransomware research. Indicative of their integration into hybrid ML pipelines, Light Gradient Boosting Machine (LightGBM) and other ensemble boosting frameworks comprise 16.00%. Last but not least, other novel and specialized models including Extreme Gradient Boosting (XGBoos), Self-Organizing Maps (SOM), Gain Ratio, and Digital DNA account for 8.00% of the overall utilization observed throughout study. Among DL methods, Recurrent NNs (RNNs) are at the forefront with 40%, followed by CNNs with 30%, CNN–RNN hybrid models at 20%, and Autoencoders with 10%. Integration of cryptographic schemes, federated learning, blockchain-based audit mechanisms, and adaptive runtime mechanisms have further boosted the mechanisms of anomaly detection with detection rates of over 99% for polymorphic and zero-day ransomware.

Author 1: Abrar Ali

Author 2: Norah Hamed

Author 3: Monir Abdullah

Keywords: Ransomware; Industrial Internet of Things; cloud computing; machine learning; deep learning; blockchain

Paper 40: A Comprehensive Survey of Visual SLAM Technology: Methods, Challenges, and Perspectives

Abstract: Visual Simultaneous Localization and Mapping (Visual SLAM) has become a cornerstone of autonomous navigation and spatial understanding in robotics, augmented reality, and computer vision. This review presents a comprehensive examination of algorithmic progress in Visual SLAM, focusing on the three principal paradigms: monocular, stereo, and RGB-D SLAM. Monocular SLAM, known for its minimal hardware requirements, has evolved from feature-based methods to deep learning-enhanced systems, addressing challenges like scale ambiguity and drift. Stereo SLAM leverages depth through triangulation, improving scale accuracy and robustness, particularly in dynamic and low-texture environments. RGB-D SLAM, utilizing depth-sensing technology, has enabled dense and semantically enriched mapping, finding significant application in indoor and real-time scenarios. Through a chronological and technical exploration of representative methods including RatSLAM, ORB-SLAM, DSO, ProSLAM, ElasticFusion, DynaSLAM, and recent hybrid and learning-based frameworks. This review identifies major milestones and architectural innovations across paradigms. A cross-paradigm analysis highlights the trade-offs in accuracy, computational efficiency, and adaptability, while also discussing emerging trends such as semantic integration, multimodal fusion, and neural implicit representations. Furthermore, the paper outlines future directions that include lifelong learning, real-time deployment on edge devices, dynamic environment adaptation, and the convergence of geometry and learning-based pipelines. Supported by a detailed taxonomy and historical evolution illustrated in visual summaries, this review serves as a foundational reference for researchers and developers aiming to understand and contribute to the advancement of Visual SLAM technologies in both academic and real-world contexts.

Author 1: Aidos Ibrayev

Author 2: Amanzhol Bektemessov

Keywords: Visual SLAM; monocular SLAM; Stereo SLAM; RGB-D SLAM; 3D mapping; pose estimation; loop closure; semantic SLAM; deep learning; sensor fusion

Paper 41: Uneven But Accelerating: AI Adoption in Higher Education

Abstract: Artificial Intelligence (AI) is increasingly recognized as a transformative force in higher education, yet adoption remains patchy and often confined to partial implementations. Using the PRISMA protocol, this study systematically reviews 74 Scopus-indexed articles published between 2015 and 2025. Publication activity rose sharply after 2020, led by contributions from China, the United States, and Saudi Arabia. Across the corpus, Perceived Usefulness and the Technology Acceptance Model (TAM) are the most frequently applied constructs, while ethical and policy dimensions remain underexamined. Thematic analysis delineates five clusters: adaptive learning and personalization; ethics and trust; digital literacy and readiness; AI in assessment and evaluation; and organizational transformation. Despite growing attention, regional gaps persist—especially in developing countries, where constrained infrastructure, funding, and digital literacy impede adoption. To address these challenges, the study proposes a multi-level conceptual framework integrating TAM, UTAUT, TPACK, and TOE to connect individual, institutional, and external factors for sustainable AI-driven education. Overall, the review underscores that AI adoption is not merely an efficiency tool but a strategic lever to advance the Sustainable Development Goals (SDGs), particularly by fostering inclusive, equitable, and innovative higher education systems.

Author 1: Mahendra Adhi Nugroho

Author 2: Umar Yeni Suyanto

Author 3: Didik Hariyanto

Author 4: Septiningdyah Arianisari

Keywords: Artificial intelligence adoption; higher education; sustainable education; developing country

Paper 42: Comparative Review of Confidence and Other Evaluation Metrics in Predictive Modeling for Procurement Fraud Coalition

Abstract: Procurement fraud, particularly when bidders act together through collusion or coalition schemes, remains a major threat to fair competition in public procurement. Predictive modeling has emerged as a key analytical tool for detecting such behaviors yet choosing appropriate evaluation metrics continues to be a challenge, especially with imbalanced or correlated data. This study applies a structured narrative review supported by a comparative analysis to examine commonly used evaluation metrics—Accuracy, Precision, Recall, F1-score, and AUC-ROC—in relation to the rule-based Confidence metric derived from association rule mining. The findings reveal that while traditional classification metrics are effective for general predictive tasks, they often fail to capture relational and co-occurrence patterns that characterize coalition fraud. In contrast, Confidence demonstrates higher interpretability and contextual relevance for detecting collusive behaviors among suppliers. The study highlights the potential of hybrid evaluation frameworks that combine classification and rule-based measures to improve fraud detection accuracy and explainability. This approach contributes to advancing predictive modeling, procurement analytics, and coalition detection by emphasizing metrics that balance performance, interpretability, and real-world applicability.

Author 1: Saifuddin Mohd

Author 2: Mohamad Taha Ijab

Keywords: Procurement fraud; predictive modeling; confidence; evaluation metrics; association rule mining; coalition detection; public sector analytics

Paper 43: Correcting Blue-Shift in Single-Image Dehazing via Haze-Compensated Von Kries Adaptation

Abstract: Haze severely degrades image quality by reducing contrast, obscuring details, and introducing a blue-shift color cast caused by atmospheric scattering. Traditional dehazing methods, including prior-based approaches (e.g., DCP, CAP, LPMinVP) and preprocessing techniques (e.g., ICAP WB, Dynamic Gamma), improve visibility but fail to correct haze-induced color imbalance, resulting in unstable RGB distributions and unnatural tone reproduction. This study proposes the Haze-Compensated Color Von Kries (HCCVK) method, a lightweight and training-free preprocessing strategy that performs color compensation before transmission estimation in single-image dehazing. HCCVK integrates a novel red-channel compensation mechanism with Von Kries chromatic adaptation to mitigate wavelength-dependent haze suppression and stabilize chromatic consistency under varying illumination. Unlike learning-based color correction approaches, HCCVK does not require training data, is computationally efficient, and maintains algorithmic interpretability, making it suitable for practical deployment. The method was evaluated on six benchmark datasets: CHIC, Dense-Haze, I-Haze, O-Haze, SOT, and NH-Haze, covering indoor, outdoor, dense, and non-homogeneous haze scenarios. Experimental results based on the RGB color balance metric (σRGB) show that HCCVK reduces color deviation by approximately 75–92% on CHIC, 80–90% on Dense-Haze, and 82–90% on NH-Haze compared to the widely used DCP, and also outperforms CAP, ICAP WB, Dynamic Gamma, and LPMinVP by producing more compact and stable RGB distributions. These findings demonstrate that HCCVK effectively corrects blue-shift imbalance, preserves luminance consistency, and enhances the color stability of dehazing pipelines.

Author 1: Asniyani Nur Haidar Abdullah

Author 2: Mohd Shafry Mohd Rahim

Author 3: Sim Hiew Moi

Author 4: Azah Kamilah Draman

Author 5: Ahmad Hoirul Basori

Author 6: Novanto Yudistira

Keywords: Image dehazing; blue-shift correction; color compensation; Von Kries adaptation; preprocessing

Paper 44: Enhancing Dermatological Diagnostics: An Enhanced Approach for Skin Cancer Classification Using pix2pix GAN

Abstract: Skin cancer is among the predominant forms of the disease that includes malignant squamous cell carcinoma, basal cell carcinoma, and melanoma that is characterized by aberrant melanocyte cell development. Frequent screenings and examinations enhance the prognosis for people with skin cancer. Sadly, a lot of patients with skin cancer are not diagnosed until the condition has progressed past the point at which treatment is effective. Deep learning techniques in computer vision have made impressive strides, but issues like class imbalance and a lack of data still hinder the autonomous identification of skin conditions. A solution to address these problems is the implementation of GAN, which is capable of synthesizing realistic data. In this paper, a deep learning GAN model for image synthesis utilizing the pix2pixHD integrated with Convolutional Neural Network (CNN) classifier approach is used to perform skin cancer classification. To categorize three forms of skin cancer benign or malignant. The proposed pix2pixHD GAN is a novel method for utilizing pertinent skin lesion information for generation of high-quality synthesized dermoscopic image and conduct skin lesion classification performance with improved accuracy. Realistic images were created using a U-Net-based generator and PatchGAN discriminator with custom CNN architecture to classify three forms of cancer. With remarkable accuracy of 87.65% (MEL), 91% (BCC), and 89.85% (SCC) and other performance parameters indicate that GAN pix2pixHD Classifier model has promising results in classification. These findings demonstrate the Classifier's ability to produce and correctly identify high-quality skin lesion images, indicating its potential as a deep learning-based medical image analysis tool.

Author 1: Adnan Afroz

Author 2: Shaheena Noor

Author 3: Shakil Ahmed Bashir

Author 4: Umair Jilani

Keywords: Deep learning; skin cancer; generative adversarial network; pix2pixHD; classification

Paper 45: A Hybrid AI Framework for DDoS Detection and Mitigation in SDN Environments Using CNN, GAN, and Semi-Supervised Learning

Abstract: The fast technological evolution seen in recent years enhanced the performance and scalability of cloud computing infrastructure and Software-Defined Networking architectures. SDN provides programmability, centralized orchestration, and dynamic resource provisioning, while separating the control and data planes to offer promising architectural paradigm for cloud computing environments. Openness and flexibility expose SDN-based networks to other security concerns, such as large-scale Distributed Denial of Service (DDoS) attacks. This paper introduces a hybrid artificial intelligence (AI) framework for detecting and mitigating DDoS attacks in SDN environments. The framework leverages three complementary approaches: Convolutional Neural Networks (CNN) to capture temporal traffic patterns, Generative Adversarial Networks (GAN) to generate synthetic traffic for dataset augmentation and to enhance anomaly detection, and semi-supervised learning techniques to exploit large amounts of unlabeled traffic data. The proposed system is deployed on a testbed combining OpenDaylight as the SDN controller and Mininet for network emulation, while the AI models are trained and run in Anaconda environment. The network traffic flows are collected, processed into statistical features (i.e., packet rates, entropy values, protocol distribution ratios), and analyzed through the hybrid AI pipeline. Mitigation actions are configured through ODL RESTCONF interface, converting the detection into OpenFlow rules to drop or rate-limit the malicious packets. Experimental evaluation demonstrates that the proposed approach achieves high accuracy detection and robustness to unseen attacks patterns demonstrating the value of applying a hybrid CNN, GAN, Semi-supervised learning approach.

Author 1: Abdelhakim HADJI

Author 2: Brahim RAOUYANE

Keywords: SDN; CNN; GAN; DDOS; OpenDaylight; Mininet; semi-supervised learning; hybrid AI framework

Paper 46: Benchmarking Deep Learning Models for Visual Classification and Segmentation of Horticultural Commodities

Abstract: Recent advances in computer vision have enabled new approaches for automated quality assessment of tropical fruits, where accurate classification and segmentation are essential for postharvest inspection. A major challenge lies in identifying deep learning architectures that achieve high accuracy while remaining computationally efficient for potential edge-based deployment. This study benchmarks three Convolutional Neural Network (CNN) models for classification (VGG16, ResNet50, and EfficientNet-B0) and two encoder–decoder models for segmentation (U-Net and DeepLabV3+) using annotated pineapple and strawberry image datasets. A 5-fold cross-validation strategy was applied to ensure statistical robustness, with evaluation metrics including accuracy, precision, recall, F1-score, Intersection over Union (IoU), and Dice coefficient. Statistical significance was verified using the Friedman and Wilcoxon signed-rank tests (α = 0.05 and 0.01). EfficientNet-B0 achieved the best classification results with average accuracies of 91.4% (strawberry) and 90.7% (pineapple), significantly outperforming ResNet50 and VGG16 (p < 0.01). For segmentation, DeepLabV3+ obtained the highest performance with mean IoU values of 91.7% and 90.8% and Dice coefficients above 92%, indicating precise boundary delineation of ripe and defective regions. Computational efficiency analysis further showed that EfficientNet-B0 had the lowest inference time (0.026 s) and smallest model size (20.4 MB), making it ideal for real-time or embedded applications. Visual analysis confirmed that DeepLabV3+ maintained robustness at fruit boundaries, though minor misclassifications were observed. This benchmarking highlights the combination of EfficientNet-B0 and DeepLabV3+ as a reliable baseline for deep learning-based fruit quality assessment.

Author 1: Fuzy Yustika Manik

Author 2: Syahril Efendi

Author 3: Jos Timanta Tarigan

Author 4: Maya Silvi Lydia

Keywords: Fruit quality assessment; classification; segmentation; EfficientNet-B0; DeepLabV3+; AISAM-CSNet

Paper 47: Towards Designing a Blockchain-Based Model for E-Book Publishing

Abstract: This paper examines the application of blockchain technology in e-book publishing by analyzing previous research and identifying current limitations. The study investigates how smart contracts and cryptographic algorithms can facilitate agreements between publishers and authors. While blockchain has been widely adopted in digital publishing domains, such as image, video, music, and scientific journals, research on its application to e-books remains limited. Existing solutions typically address individual challenges such as transaction transparency, authenticity, or copyright protection, but rarely integrate them into a single framework. To provide a systematic synthesis of prior works, this paper develops a taxonomy of blockchain-based e-book publishing models across six dimensions: platform, storage, smart contract usage, cryptographic algorithm, tokenization, and actors. This paper reviews seven (7) blockchain-based models in e-book publishing and identifies their limitations. Based on these insights, a conceptual blockchain-based smart contract model for e-book publishing was proposed using Ethereum platform, incorporating InterPlanetary File System (IPFS) storage and cryptographic algorithms. The proposed model has the potential to significantly enhance the security and rights protection for authors and publishers, thereby fostering a more secure and equitable e-book publishing landscape.

Author 1: Maznun Arifa Mohammadan Makhtar

Author 2: Novia Admodisastro

Author 3: Suleymenova Laura Askarbekkyzy

Keywords: e-book publishing; Ethereum; blockchain technology; smart contract

Paper 48: Adaptive Hybrid Deep Learning with Recursive Feature Elimination for Physical Violence Detection

Abstract: Physical violence among students remains a persistent issue that often goes undetected, especially in school environments without intelligent real-time monitoring systems. Such incidents pose serious risks to student safety and hinder the creation of a secure learning atmosphere. This study aims to develop an adaptive visual-based system for detecting physical violence in educational settings using a deep learning approach. A hybrid architecture was designed by integrating VGG19 for spatial feature extraction and Bidirectional Long Short-Term Memory (BiLSTM) for temporal sequence analysis. To enhance model interpretability and reduce redundancy, Recursive Feature Elimination (RFE) was employed to eliminate irrelevant features and improve overall learning efficiency. The proposed system effectively captures both spatial and temporal cues from classroom surveillance videos, enabling more accurate classification of violent and non-violent behaviors. The model was trained and tested on benchmark datasets containing diverse video samples and achieved an accuracy of 92.4%, outperforming standalone CNN and LSTM models. The integration of RFE contributed to a more compact and computationally efficient framework. This study demonstrates the potential of hybrid deep learning and feature optimization for real-time violence detection, contributing to the advancement of visual intelligence and Educational AI for safer, data-driven learning environments.

Author 1: Sukmawati Anggraeni Putri

Author 2: Duwi Cahya Putri Buani

Author 3: Achmad Rifa’i

Author 4: Imam Nawawi

Keywords: Violence detection; deep learning; VGG19; BiLSTM; RFE; Educational AI

Paper 49: Transformative Integration of Machine Learning in Software Applications in Light of Current Software Engineering Practices

Abstract: This study critically reviews the transformative integration of machine learning (ML) into software engineering, detailing its evolution from traditional DevOps to MLOps, which has significantly enhanced software development by enabling adaptive and intelligent systems, improving processes, and boosting software quality. Despite these benefits, the integration introduces unique challenges across technical (e.g., model deployment, data quality, scalability), organizational (e.g., collaboration, tool management), and cultural (e.g., resistance to change, skill gaps) domains throughout the software development lifecycle. The review highlights emerging solutions, including robust MLOps practices, microservices architecture, and frameworks like CRISP-DM, DataOps, and Agile ML, which aim to streamline the ML lifecycle and ensure reliability and scalability. Furthermore, it emphasizes the crucial role of security and governance frameworks in protecting against adversarial attacks, maintaining data privacy, and ensuring accountability and compliance, which are essential for building trust and ethical application of ML systems. Ultimately, successful ML integration requires a holistic approach that addresses these multifaceted challenges to optimize ML's impact and drive technological progress and business value.

Author 1: Fawzi Abdulaziz Albalooshi

Keywords: Machine learning (ML); software engineering; DevOps; MLOps; ML integration challenges; integrated software development

Paper 50: Bridging Machine-Readable Code of Regulations and its Application on Generative AI: A Survey

Abstract: Machine-Readable Code (MRC) and Machine-Readable Regulations (MRR) enable the conversion of complex regulations into structured formats such as JSON, XML, and X2RL, allowing machines to parse and interpret regulatory texts efficiently. Currently, organizations face challenges in regulatory compliance due to the complexity of regulations, frequent updates, and difficulty in identifying changes that impact policies and procedures. Existing literature provides guidance to a certain extent on how to anticipate regulatory modifications or ensure timely compliance. This review examines current literature on applying machine learning (ML) and Generative AI (GenAI) to extract, structure, and interpret regulatory content. It surveys techniques for converting regulations into machine-readable formats, predicting regulatory changes, and assessing alignment with real-world modifications issued by regulatory bodies. The findings indicate that using MRC, MRR, and AI enables automated compliance checks, faster detection of violations or errors, standardized compliance processes, real-time monitoring, and automatic report generation. These approaches can significantly enhance regulatory adherence across industries, particularly in sectors such as finance, where compliance is critical.

Author 1: Samira Yeasmin

Author 2: Bader Alshemaimri

Keywords: Regulatory compliance; natural language processing; machine learning; machine-readable code; Machine-Readable Regulations; generative AI; large language models; RegTech; conflicting regulations; regulation issuance

Paper 51: Energy Efficient Workflow Allocation in Cloud Computing Using Improved Grey Wolf Optimization

Abstract: Cloud computing has emerged as a dominant platform for hosting complex applications, offering scalable and flexible resources on demand. However, the dynamic and heterogeneous nature of cloud environments poses significant challenges for efficient workflow scheduling, particularly when aiming to minimize total execution time, energy consumption, and operational cost. In this research, we propose a novel hybrid approach that integrates the Heterogeneous Earliest Finish Time (HEFT) algorithm with an Improved Grey Wolf Optimizer (IGWO) enhanced by differential evolution strategies and survival-of-the-fittest mechanisms. These enhancements strengthen exploration and exploitation by adaptively mutating and refining task allocations while eliminating weaker solutions. The use of HEFT-based initialization provides a strong starting population, and the DE-driven IGWO refinement accelerates convergence and avoids premature stagnation. Together, these two-level optimization strategy ensures faster convergence and higher energy-efficient workflow scheduling compared to earlier HEFT metaheuristic approaches. To evaluate the effectiveness of the proposed hybrid method, extensive experiments were conducted on randomly generated workflows with varying task and dependency complexities. The performance analysis demonstrates that the hybrid HEFT-IGWO approach consistently outperforms standard HEFT, traditional GWO, and standalone metaheuristic techniques in terms of minimizing makespan, reducing energy consumption, and lowering cloud infrastructure costs. This study highlights the potential of combining heuristic initialization with evolutionary optimization to achieve energy-efficient, cost-effective workflow scheduling in cloud computing environments.

Author 1: Md. Mazhar Nezami

Author 2: Anoop Kumar

Keywords: Cloud computing; energy efficient; workflow; Heterogeneous Earliest Finish Time (HEFT); Grey Wolf Optimization (GWO); makespan; cost

Paper 52: Feature Pyramid Network with Dual-Decoder Supervision for Accurate Stroke Lesion Localization in Multi-Modal Brain MRI

Abstract: This study presents a novel Feature Pyramid Network with Dual-Decoder Supervision for accurate stroke lesion localization in multi-modal brain MRI. The proposed architecture integrates a Swin Transformer backbone with multi-scale feature aggregation, enabling effective fusion of hierarchical representations from DWI, ADC, and FLAIR sequences. A dual-decoder structure is employed, where the auxiliary decoder provides coarse lesion guidance through pseudo masks, and the primary decoder refines boundaries for precise voxel-level segmentation. Auxiliary supervision improves convergence stability and feature discrimination, while modality dropout enhances robustness to incomplete imaging protocols. Experiments conducted on the ATLAS v2.0 dataset demonstrate superior performance over baseline encoder–decoder models, achieving higher Dice scores, improved boundary accuracy, and strong lesion-wise detection rates. The model consistently localizes lesions of varying size, shape, and intensity, with minimal overfitting, as evidenced by small training–testing performance gaps. Qualitative results confirm the framework’s ability to transform coarse localization into anatomically accurate predictions. The combination of multi-modal integration, dual-decoder specialization, and self-training mechanisms positions the proposed method as a promising candidate for clinical deployment in rapid stroke diagnosis workflows. Future directions include expanding validation to multi-center datasets, incorporating explainable AI techniques, and enabling real-time 3D processing for deployment in acute care environments.

Author 1: Satmyrza Mamikov

Author 2: Zhansaya Yakhiya

Author 3: Bauyrzhan Omarov

Author 4: Yernar Mamashov

Author 5: Akbayan Aliyeva

Author 6: Balzhan Tursynbek

Keywords: Stroke lesion localization; multi-modal MRI; feature pyramid network; segmentation; deep learning

Paper 53: Leveraging AI and Hybrid Intelligence for Robust Geospatial Data Fusion in Autonomous Terrestrial Navigation

Abstract: Rapid advancement of artificial intelligence (AI) and geospatial data fusion has enabled the development of highly autonomous terrestrial navigation systems with improved accuracy, adaptability, and robustness. This paper proposes a novel framework integrating multi-source geospatial data fusion with deep learning-based decision-making for autonomous terrestrial navigation. Unlike conventional approaches that rely solely on Global Navigation Satellite Systems (GNSS) or inertial sensors, our system leverages a hybrid fusion model combining GNSS, LiDAR, camera vision, and high-resolution geospatial databases. A deep reinforcement learning (DRL) paradigm is introduced to enhance the system’s adaptability in dynamic environments, optimizing route planning and obstacle avoidance in real-time. Additionally, a hybrid AI model incorporating Graph Neural Networks (GNN) and Transformer-based architectures processes spatial and temporal dependencies in navigation data, improving localization precision and resilience against sensor failures. The proposed system is evaluated through extensive simulations and real-world tests, demonstrating superior performance in complex urban and off-road scenarios compared to traditional Kalman filter-based methods. Our findings highlight the potential of AI-driven geospatial data fusion in redefining autonomous navigation, paving the way for next-generation intelligent mobility solutions.

Author 1: Manel Salhi

Author 2: Mounir Bouzguenda

Author 3: Faouzi Benzarti

Author 4: Fawaz Alanazi

Author 5: Ezzeddine Touti

Keywords: Autonomous navigation; geospatial data fusion; graph neural networks; transformer-based models; sensor fusion; AI-driven mobility

Paper 54: SEARCHX: An Integrated Framework of Distributed Intelligent Search Services Based on Web Browser

Abstract: To address the growing demand for web search and improve the performance and accuracy of search systems, this study proposes a distributed intelligent search service integration framework based on SEARCHX. This framework leverages the local computational power of the browser, integrating inverted indexing, data sharding, and replication mechanisms, as well as the Term Frequency-Inverse Document Frequency (TF-IDF) intelligent ranking algorithm. These components enable front-end distributed processing of search tasks and multi-source result fusion. Experiments are conducted on six major browser platforms, Chrome, Firefox, Edge, Safari, etc., using the open-source Text REtrieval Conference (TREC) dataset. The system’s response performance and accuracy are evaluated under varying search loads. The experimental results show that, compared to the unoptimized version, the optimized SEARCHX reduces the average response time by approximately 27 per cent under medium-to-high load conditions. Precision improves by an average of 0.05, and the F1 score increased by more than 0.04 on all platforms. The system also demonstrates good stability and consistency across multiple platforms. SEARCHX provides a viable approach to building decentralized, high-efficiency, and easily deployable intelligent search services, with strong practical value and expansion potential. This study aims to construct a decentralized, cross-platform, and high-performance intelligent search service framework, offering a more efficient, stable, and accurate technical support solution for users in complex search environments.

Author 1: Zehui Zhang

Author 2: Lin Zhou

Author 3: Jie Peng

Author 4: Liwei Wang

Author 5: Bo Cheng

Keywords: Web; TF-IDF; distributed network; intelligent search; SEARCHX

Paper 55: Multi-Criteria Using Dijkstra’s Algorithm to Determine Optimal Time Paths in Vehicle Route Optimization

Abstract: This research is the development of a Road Traffic Network using one of the methods in Mathematics, namely Dijkstra's Algorithm, Weighted Sum Method (WSM), and Weighted Product Method (WPM). Meanwhile, the parameters used are route, volume, capacity, DOS, distance, and travel time. The objective of this research is to find the fastest route alternative from one place to another. The recommended results using Dijkstra's algorithm combine values of distance, travel time, and congestion degree with the weighted-sum and weighted-product methods, each calculated accordingly. The shortest route is 1→2→4→7→10→14, the route with the least congestion and shortest travel time is 1→2→5→6→9→13→14. This research combines these three parameters to obtain a balanced route between congestion level, short travel distance, and short travel time for the driver. By combining these parameters, the best route from this study is route 1: 1→2→5→8→12→11→14 with a total distance of 28.77 km, a saturation degree value of 5.421, and a travel time of 28 minutes. Thus, the research results indicate that the best route will have a combination of multiple criteria, such as short distance, short travel time, and less congestion simultaneously. The Weighted-Sum Method (WSM) and Weighted-Product Method (WPM) can produce different outputs, with WPM being superior to WSM in terms of computational steps.

Author 1: Basorudin

Author 2: Handaru Jati

Author 3: Nurkhamid

Author 4: Puput Dani Prasetyo Adi

Keywords: Dijkstra’s algorithm; multi-criteria; road traffic network; Weighted Sum Method (WSM); Weighted Product Method (WPM); mathematics

Paper 56: Embedding Models: A Comprehensive Review with Task-Oriented Assessment

Abstract: Sentence embedding is a very important technique in most natural language processing (NLP) tasks, such as answer generation, semantic similarity detection, text classification and information retrieval. This technique aims to transform the semantic meaning of a sentence into a fixed-dimensional vector, allowing machines to understand human language. Sentence embedding has moved in recent years from simple word vector averaging methods to the development of more sophisticated models, particularly those based on transformer structures such as the BERT model and its variants. However, systematic reviews that critical, analyze and compare the performance of these models are still limited, particularly the selection of the appropriate embedding model for a specific NLP task. This study aims to address this gap by a comprehensive review for sentence embedding models and a systematic evaluation of their performance on NLP tasks, such as semantic similarity, clustering, and retrieval. The study enabled us to identify the appropriate embedding model for each task, identify the main challenges faced by embedding models, and propose effective solutions to improve the performance and efficiency of sentence embedding.

Author 1: Lahbib Ajallouda

Author 2: Meriem Hassani Saissi

Author 3: Ahmed Zellou

Keywords: Natural language processing; sentence embedding models; transformer models; embedding models challenges

Paper 57: Machine Learning-Driven Emotional Feedback Analysis and Adaptive Content Generation for VR Movie and TV Users

Abstract: With the growing demand for immersive audiovisual experiences, user sentiment feedback analysis has become a pivotal factor in improving personalization and interactivity in virtual reality (VR) movie and television. This study proposes a machine learning–driven framework that integrates sentiment feedback recognition and adaptive content generation to optimize user experience. First, a Long Short-Term Memory (LSTM) model is developed to analyze multimodal sentiment feedback data, including physiological signals, behavioral responses, and interactive actions. The model achieves an average recognition accuracy of 75.75% across four basic emotions—happiness, sadness, anger, and fear—demonstrating its ability to capture dynamic and continuous emotional patterns. Based on real-time sentiment feedback, a Deep Q-Network (DQN) reinforcement learning algorithm is employed to generate adaptive VR content that aligns with users’ current emotional states. Experimental validation with 100 participants shows that adaptive content generation increases overall satisfaction scores from 6.2 to 7.8, and the matching degree between user emotions and content improves by more than 20%. The integration of sentiment feedback analysis and reinforcement learning establishes a closed feedback loop—emotion detection → adaptive adjustment → feedback optimization—that enhances immersion, empathy, and user engagement. This research provides a data-driven reference for the intelligent evolution of VR movie and television, and future work will expand to fine-grained emotional dimensions and multimodal fusion to improve recognition precision and real-time adaptive generation performance.

Author 1: Yun TANG

Keywords: Machine learning; VR movie and television; user sentiment feedback analysis; adaptive content generation; reinforcement learning

Paper 58: Meta-Learning Prediction Framework for Asphalt Mixtures Fatigue Life Modeling

Abstract: In order to improve the accuracy and generalization ability of asphalt mixture fatigue life prediction, this study introduces the meta-learning method, which aims to solve the problems of poor adaptability and strong data dependence of the traditional prediction model under complex working conditions. In this study, a prediction framework based on the Model-Agnostic Meta-Learning (MAML) algorithm is constructed, which realizes the fast and accurate prediction of asphalt mixture fatigue life under multi-task conditions through feature extraction, meta-knowledge learning, and a fast adaptive mechanism. The experiments were conducted using multi-class mixture data and compared with linear regression and BP neural network methods under the MATLAB platform. The results show that the meta-learning model achieves a prediction accuracy of 0.98 within 500 iterations, which is significantly better than that of the BP neural network (0.89) and linear regression (0.84), and the prediction error is controlled to be between 40 and 60 under typical working conditions, while the traditional method has an error of up to 150. Further analysis shows that the meta-learning method has a faster convergence rate (the convergence index is 0.9 for 100 iterations) and a higher convergence index of 0.9 for 100 iterations. 0.9) with higher robustness. In conclusion, the meta-learning-based prediction method shows excellent performance in fatigue life modeling, which is suitable for rapid application in real-world engineering with diverse materials and loading environments.

Author 1: Longmeng Tan

Author 2: Krzysztof Kowalski

Keywords: Asphalt mixtures; fatigue life; meta-learning prediction; mechanism analysis

Paper 59: Quality Classification of Harumanis Mango Based on External Multi-Parameter and Machine Learning Techniques

Abstract: Grading Harumanis mangoes is traditionally done through manual visual inspection, which is subjective, inconsistent, and labor-intensive. Industry practices report only 70–80% consistency among human graders, with accuracy further declining under fatigue or high volumes. These limitations hinder uniform quality assurance, especially for export markets. To address this, an image-based, non-destructive grading system was developed, focusing on external features such as surface defect severity, ripeness index, shape uniformity, and size. A dataset of 1,018 mango samples was collected and analyzed using a machine vision system. Features were extracted through image segmentation and color–shape analysis, then classified using a Fuzzy Inference System (FIS) and Machine Learning (ML) models including SVM, MLPNN, and ANFIS. Enhanced SVM variants were also implemented to assess performance gains. Results showed strong performance across all parameters: ripeness index accuracy reached 93.5%, shape uniformity 91.6%, and size classification over 96%. The enhanced SVM+ achieved the best overall accuracy at 95.1% with the lowest error rates. The proposed system demonstrated clear improvements over manual grading and effectively classified mangoes into PREMIUM, GRADE 1, GRADE 2, and REJECT categories, supporting its potential for reliable real-world deployment.

Author 1: Mohd Nazri Abu Bakar

Author 2: Abu Hassan Abdullah

Author 3: Muhamad Imran Ahmad

Author 4: Norasmadi Abdul Rahim

Author 5: Haniza Yazid

Author 6: Wan Mohd Faizal Wan Nik

Author 7: Shafie Omar

Author 8: Shahrul Fazly Man@Sulaiman

Author 9: Tan Shie Chow

Author 10: Fahmy Rinanda Saputri

Keywords: Machine learning; image processing; quality assessment; Harumanis mango; appearance attributes

Paper 60: Fourier Transform and Attention Guided Deep Neural Network for Face Anti-Spoofing in Medical Applications

Abstract: Face recognition systems have become prevalent in mobile devices and security applications, increasing the demand for robust face presentation attack detection. Early efforts based on handcrafted features struggled to cope with variations in illumination, pose, and attack modalities, prompting a transition toward deep learning solutions capable of extracting subtle discriminative cues. A novel architecture built upon an EfficientNet-V2 backbone, combined with a Shuffle Attention module and Fourier heads, was developed to capture both spatial and frequency domain characteristics. A dual-path approach processes each input face image through conventional convolutional blocks and a 2D Discrete Fourier Transform path, with dedicated Fourier heads reconstructing frequency maps that reveal minute discrepancies between genuine and spoofed presentations. Experimental evaluation on the Oulu-NPU dataset demonstrates strong performance across four protocols, including robust detection under varying environmental conditions, low error rates with novel attack types, and consistent results across different sensor inputs. Metrics such as APCER, BPCER, and ACER validate the method’s ability to distinguish between live and fake faces reliably. The outcomes suggest that combining spatial and frequency cues addresses limitations observed in earlier approaches, offering valuable insights for deployment in security-sensitive applications and setting a strong foundation for future research in face anti-spoofing.

Author 1: Zhanseri Ikram

Keywords: Liveness detection; face anti-spoofing; deep learning; CNN; frequency domain

Paper 61: A Technique for Automated Parallel Optimization of Function Calls in C++ Code

Abstract: In modern software development, achieving high performance increasingly relies on effective parallelization. While much of the existing research has focused on loop-level parallelism, function-level parallelization remains relatively underutilized. Yet, in many real-world applications, function calls serve as natural units of computation that could greatly benefit from concurrent execution. To address this gap, we present an automated tool that analyzes sequential C++ code, identifies independent function calls, and evaluates their suitability for parallel execution. The tool performs three key analyses: dependency analysis to detect function calls, context analysis to understand execution conditions, and workload assessment to determine whether parallelization would result in significant performance benefits. Based on the analysis results, the tool transforms eligible function calls into parallel equivalents without altering the original program logic. Additionally, the tool generates detailed Control Flow Graphs (CFG) for each function in three formats, facilitating further structural analysis. Three benchmark programs were used in experimental testing. The evaluation measured both sequential and parallel execution times, along with the computed performance gain expressed as a percentage reduction in runtime. Results demonstrated the tool’s ability to improve execution efficiency and reduce processing time. These outcomes emphasize the tool’s role in advancing function-level automatic parallelization. The tool showed notable performance improvements across the three benchmark applications, with the Employee Performance System achieving the highest improvement of 54.6%, followed by the Genomic Sequence System at 48.3%, and the Book Reviews System achieving an improvement of 36.1%. Demonstrating the tool’s ability to improve efficiency via automated function-level parallelization.

Author 1: Shuruq Abed Alsaedi

Author 2: Fathy Elbouraey Eassa

Author 3: Amal Abdullah AlMansour

Author 4: Lama Abdulaziz Al Khuzayem

Author 5: Rsha Talal Mirza

Keywords: Automatic parallelization; function-level parallelization; C++ code optimization; parallel computing; control flow graph; dependency analysis; performance optimization

Paper 62: Unveiling the Drivers of Consumer Purchase Intention in Short-Form Video Marketing

Abstract: Short-video features on e-commerce platforms have become a key driver of social commerce, enhancing user engagement and purchase intention. However, user reviews of Shopee Video reveal issues such as disruptive autoplay, limited content control, and unintuitive navigation. While prior studies have examined engagement and satisfaction in general e-commerce, limited research has explored how short-video features within social commerce influence purchase intention through user engagement. This study fills that gap by analysing the factors affecting consumer behaviour toward Shopee Video. Sentiment analysis of user reviews identified common dissatisfaction themes, followed by a quantitative survey of 300 Shopee Video users in Indonesia. Using Structural Equation Modeling with the Partial Least Squares (SEM-PLS) approach, the results show that user engagement significantly mediates the relationship between system quality, information quality, and technology adoption and usefulness on purchase intention. The study extends existing models to the social commerce context, providing insights for optimizing short-video features to strengthen engagement and conversion.

Author 1: Merisa Syafrina

Author 2: Viany Utami Tjhin

Keywords: E-commerce; purchase intention; Shopee Video; social commerce; SEM-PLS

Paper 63: Integrative Hybrid Metaheuristic Algorithm for Hyperparameter Optimisation in Pre-Trained Convolutional Neural Network Models (I-HAHO)

Abstract: Hyperparameter optimisation (HPO) remains a fundamental challenge in deep learning, especially for pre-trained convolutional neural networks (CNNs). While pre-trained models reduce the computational burden of training from scratch, their effectiveness depends heavily on tuning parameters such as learning rate, batch size, dropout, weight decay, and optimizer type. The search space of hyperparameters is large, nonlinear, and highly dataset-dependent, making traditional techniques like grid search, random search, and Bayesian optimisation insufficient. This paper introduces I-HAHO, an Integrative Hybrid Metaheuristic Algorithm that combines Artificial Bee Colony (ABC) for global exploration and Harris Hawks Optimisation (HHO) for local exploitation. A diversity-based phase-switching mechanism dynamically regulates exploration and exploitation, allowing the optimiser to adapt its search behaviour to varying landscape conditions. Experiments on CIFAR-10, CIFAR-100, SVHN, and TinyImageNet with three CNN architectures (VGG16, ResNet50, EfficientNet-B0) demonstrate up to 6.9% accuracy improvements. I-HAHO enhances adaptability, scalability, and robustness for hyperparameter tuning.

Author 1: Nazleeni Samiha Haron

Author 2: Jafreezal Jaafar

Author 3: Izzatdin Abdul Aziz

Author 4: Mohd Hilmi Hasan

Author 5: Muhammad Hamza Azam

Keywords: Hyperparameter Optimisation (HPO); Convolutional Neural Networks (CNNs); Artificial Bee Colony (ABC); Harris Hawks Optimisation (HHO); Hybrid Metaheuristic Algorithm

Paper 64: A Novel Performance-Based Time Series Forecast Combination Method and Applications with Neural Networks

Abstract: Performance-based forecast combination approaches determine the weights of the individual forecasts based on the inverse average error for a past time interval. However, although the performances are calculated for a time span, the aim is mostly a one-step-ahead time-point forecast. In these classical methods, a relatively higher prediction error of a single past time-point spreads and decreases the performance value of the model, even though the model is highly successful on other time-points in the interval. In this study, a novel approach is presented where performance of each past time-point prediction is calculated separately. Instead of taking the inverse average error for a pre-determined past time interval, prediction performance is calculated for each past data point separately using the normalized inverse absolute error, then the average performances are calculated for past time interval to get the combination weights. To be able to measure the performance of the presented methodology, it is applied on three well-known time series data. Seven different models of neural networks, based on multi-layer perceptron and extreme learning machines are used to model, forecast and form the combination forecasts. Moreover, four different performance-based combination techniques, two central tendency-based benchmark combination methods and the naïve model are employed for comparison. The obtained results show that proposed methodology is a powerful and robust technique and superior to all performance-based combination techniques compared.

Author 1: M. Burak Erturan

Keywords: Combination forecast; performance-based combination; neural networks; multi-layer perceptron; extreme learning machine

Paper 65: Integration of Color QR-Code Technology in Biometric Data Encoding and Facial Identity Systems

Abstract: This paper presents an enhanced algorithm for the generation of color biometric QR codes capable of encoding facial image data, anthropometric parameters, and personal identity information simultaneously within a single RGB-based QR structure. The proposed approach extends existing monochrome QR models by integrating optimized image decomposition, modular QR block generation, and multi-channel RGB encoding to achieve higher data density, improved privacy protection, and better readability under various lighting and compression conditions. The algorithm was implemented in Python using the OpenCV library, ensuring compatibility with contemporary biometric systems, embedded devices, and mobile platforms. Experimental evaluations conducted on standard face databases demonstrate the method’s robustness in terms of decoding accuracy, distortion resilience, and information integrity. Furthermore, the study explores new applications such as animated QR codes and photo–sketch hybrid datasets for training and validation purposes. The results highlight the potential of color biometric QR technology for secure identification, access control, and digital identity verification, offering a novel bridge between computer vision and information security.

Author 1: Nazym Kaziyeva

Author 2: Kalybek Maulenov

Author 3: Ruslan Ospanov

Author 4: Abzhan Khamza Mukhtaruly

Keywords: Color biometric QR code; facial image encoding; RGB channel decomposition; biometric data integration; secure identification; facial recognition; QR animation; identity encoding; privacy protection; data capacity; OpenCV; computer vision

Paper 66: Enhancing the Scanability of Damaged QR Codes Through Image Restoration Using GANs Combined with the Spectral Normalization Technique

Abstract: QR Codes are widely used in the digital era for storing and sharing information in various applications. However, they are often susceptible to physical damage such as scratches, tears, or fading, which can result in scanning failures and limit their usability. To overcome this issue, this research introduces a Generative Adversarial Network (GAN) model integrated with Spectral Normalization to restore damaged QR Code images. The model was trained and evaluated using a dataset of QR Codes with simulated damage ranging from 1% to 60%. Experimental results demonstrate that the proposed approach effectively reconstructs missing parts of QR Codes while preserving structural details and module sharpness. The model achieved an average PSNR of 28.5 dB, SSIM of 0.91, and a scanning success rate of 88%, outperforming U-Net (68%) and a baseline GAN (75%). Although the processing time is slightly longer, the model offers superior accuracy and robustness, particularly for severely damaged QR Codes (40% to 60% damage). These findings confirm that GANs enhanced with Spectral Normalization offer a promising solution for QR Code restoration, with potential uses in digital marketing, payment systems, and inventory management.

Author 1: Puwadol Sirikongtham

Author 2: Apichaya Nimkoompai

Keywords: QR Code restoration; Generative Adversarial Networks; spectral normalization; image inpainting; deep learning; damage reconstruction

Paper 67: Recent Integrating Machine Learning and Malay-Arabic Lexical Mapping for Halal Food Classification

Abstract: The rapid growth of e-commerce has changed the way people engage with businesses, notably in the food industry. For the Muslim community, guaranteeing Halal conformity in digital transactions is critical. This study provides a comprehensive framework for improving Halal E-Commerce systems that include machine learning, pattern libraries, and multilingual support, specifically in Malay and Arabic. The study examines the role of pattern libraries in designing user-friendly interfaces, as well as lexical mapping strategies for enhancing Malay-Arabic translation accuracy. Natural language processing (NLP) and machine learning are combined to create an application that classifies food items into two categories: Halal or Haram. With an accuracy of 85%, a Random Forest classifier is trained on labeled datasets. Preparing the text, extracting features using TF-IDF, and evaluating the results using precision, recall, and F1-score are all steps in the classification process. To increase classification accuracy, a rule-based approach is also applied to conditional logic and keyword matching. By adjusting the parameters, the model is further improved, leading to strong performance. By taking into account the cultural and linguistic requirements of the Muslim community, multilingual support enhances accessibility and user confidence. The suggested method increases translation accuracy by employing lexical mapping at the word, phrase, and context levels. The paper also assesses several machine learning models, demonstrating that Random Forest outperforms the other methods examined. The findings contribute to the growth of Halal E-Commerce by outlining a systematic strategy to ensure compliance and usability. The proposed system can serve as a platform for future research into AI-driven Halal certification and digital marketplace optimization, blockchain with an e-Commerce framework.

Author 1: Noorrezam Yusop

Author 2: Massila Kamalrudin

Author 3: Nuridawati Mustafa

Author 4: Tao Hai

Author 5: Mohd Nazrien Zaraini

Author 6: Halimaton Hakimi

Author 7: Siti Fairuz Nurr Sardikan

Keywords: Malay-Arabic lexical mapping; natural language processing; machine learning; halal food classification; halal food e-commerce

Paper 68: Evaluating Head Pose Estimation for Assessing Visual Attention in Children with Special Needs During Robot-Assisted Therapy

Abstract: This study investigates the application of head pose estimation (HPE) to assess visual attention in children with special needs (CwSN) during robot-assisted therapy sessions, focusing on its effectiveness and the attention patterns exhibited by these children. CwSN often faces unique challenges, such as sensory processing difficulties or delayed cognitive processing. Age and therapy duration also influenced attention levels, with younger children generally exhibiting shorter attention spans than older participants. Additionally, familiarity with technology, such as prior screen time at home, positively impacted engagement during robot-assisted therapy. An experimental study was conducted with 30 children aged 2 to 7 years, including those with autism spectrum disorder (ASD), speech delay (SD), and attention-deficit/hyperactivity disorder (ADHD). Using an integrated camera, head movements were tracked to analyse forward-facing head direction as an indicator of attention. The system achieved an overall accuracy of 82% and an average attention percentage of 65%, highlighting that visual attention varies significantly based on the type of disability, age, and therapy duration. The integration of the robot enhanced visual engagement across all groups, fostering improved interaction and attention. These findings emphasise the importance of tailoring robot-assisted therapy (RAT) to the specific needs and attention patterns of children with different disabilities, ages, and therapy histories, underscoring the potential of assistive robotics to optimise therapeutic outcomes in special education settings. This research highlights the potential of personalised RAT to improve social, cognitive, and motor skills. It offers evidence-based strategies for integrating assistive robotics into special education and therapeutic settings for CwSN.

Author 1: Rusnani Yahya

Author 2: Rozita Jailani

Author 3: Nur Khalidah Zakaria

Author 4: Fazah Akhtar Hanapiah

Keywords: Head pose estimation; visual attention; robot-assisted therapy; children with special needs

Paper 69: A Novel Taxonomy for Human Activity Recognition Based on a Systematic Analysis of Public UAV Datasets

Abstract: In recent decades, unmanned aerial vehicles (UAVs) have become widely utilized for many real-world applications, including surveillance, crowd management, and threat detection, providing a new perspective to recognize human behaviors. However, current UAV-based video datasets adopt categorization schemes that rely on broad and inconsistent categories relative to real-world aerial contexts. To address this knowledge gap, this study proposes a novel human activity categorization framework derived from a comprehensive systematic analysis study of ten publicly available UAV-based human action recognition (HAR) datasets, incorporating a variety of environmental situations and human behaviors. By reconciling inconsistent categories and finer activities, this taxonomy serves as a standard framework for UAV-based HAR research. The proposed categorization framework is validated by comparing it with other existing frameworks on the publicly benchmarked Drone-Action dataset, outperforming them by 97% across four metrics. Our contribution aims to develop the foundation for further experimental validation and provide a guide for researchers interested in developing accurate and context-aware surveillance systems.

Author 1: Sumaya Abdulrahman Altuwairqi

Author 2: Salma Kammoun Jarraya

Keywords: Human action recognition; UAV videos; surveillance systems; categorization framework

Paper 70: Evaluation of the Impact of Cybersecurity Knowledge on the Prevention of Social Cybercrime Among University Students in Mexico, Colombia, and Peru

Abstract: Objectives: This study aims to evaluate the degree of cybersecurity knowledge and awareness among university students in Peru, Mexico, and Colombia, and to determine how these factors contribute to protection against social cybercrime. This cross-regional analysis represents a novel contribution by comparing cybersecurity preparedness across three Latin American countries, an underrepresented region in cybersecurity education research. Methods: A cross-sectional study was conducted using a 97-question survey that assessed both cybersecurity knowledge and practices. The study involved 809 university students from Peru, Mexico, and Colombia. Correlation analysis was performed to examine the relationship between cybersecurity knowledge and cybercrime prevention practices. Results: The analysis revealed a positive but low correlation (r=0.252) between cybersecurity knowledge and cybercrime prevention practices. Only 10.71% of preventive practices could be explained by acquired knowledge. Greater efficacy was observed in cyberstalking prevention compared to other forms of cybercrime. A significant gap was found between theoretical knowledge and practical application of cybersecurity, with only 44.6% of students receiving occasional information on the subject. Conclusions: This study highlights the urgent need to improve cybersecurity education in Latin American universities. The findings underscore the importance of integrating applied practices into cybersecurity curricula to strengthen students' ability to effectively counter cyber threats. Future educational initiatives should focus on bridging the gap between theoretical knowledge and practical application to enhance students' resilience against social cybercrime.

Author 1: Yasmina Riega-Viru

Author 2: Lainiver Mendoza Munar

Author 3: Mario Ninaquispe-Soto

Author 4: Kiara Nilupu-Moreno

Author 5: Juan Luis Salas-Riega

Author 6: Alfonso Renato Vargas-Murillo

Author 7: Yolanda Pinto Bouroncle

Keywords: Cybersecurity; social cybercrime; university students

Paper 71: Understanding Echo Chambers in Recommender Systems: A Systematic Review

Abstract: Echo chambers refer to the phenomenon in which individuals are consistently exposed to content that aligns with their existing viewpoints. Over time, this can narrow a user’s perspective and make it harder to encounter different opinions. In this systematic literature review, we looked at studies published between 2019 and early 2025 and how they have approached this issue, from understanding the cause, and examining existing detection and mitigation strategies. We went through and organized the main findings, noting patterns in the algorithms used, the role of user behavior, and the influence of the data itself. Several works also suggest ways to introduce more variety into recommendations, aiming to break repetitive exposure. Our review confirms that echo chambers and filter bubbles do exist in recommender systems and that they raise concerns for diversity and fairness. Furthermore, we end by pointing to open questions and possible directions for future work, for both researchers and practitioners.

Author 1: Meriem HASSANI SAISSI

Author 2: Nouhaila IDRISSI

Author 3: Ahmed ZELLOU

Keywords: Echo chamber; recommender systems; filter bubbles; collaborative filtering; systematic literature review

Paper 72: Critical Review of Object Detection Techniques for Traffic Light Detection in Intelligent Transportation Systems

Abstract: Object detection and tracking play a critical role in intelligent transportation systems (ITS), particularly in recognizing and monitoring traffic lights to ensure safety and improve traffic efficiency. Despite progress in deep learning and optimization algorithms, traffic light detection still faces persistent challenges under varying conditions such as illumination changes, occlusions, and visual clutter. This study provides a critical review of object detection techniques specifically for traffic light detection, evaluating the evolution of machine learning frameworks, deep learning architectures, and hybrid optimization models. The review identifies research gaps in the robustness, real-time adaptability, and generalizability of existing methods. Furthermore, it highlights emerging trends such as multi-camera systems, anchor-free detection, and hybrid optimization techniques that bridge performance trade-offs between accuracy and efficiency. The findings offer a new perspective on integrating multiple approaches to achieve scalable, high-accuracy traffic light detection for future ITS applications.

Author 1: Adhwa Salemi

Author 2: Muhammad Arif Mohamad

Keywords: Object detection; traffic light detection; optimization; intelligent transportation systems; review

Paper 73: Conversational AI-Powered VR Development Model for Tourism Promotion in Thailand: Expert Assessment and Stakeholder Acceptance

Abstract: Thailand’s tourism sector increasingly requires immersive digital innovations that preserve local identity while enhancing visitor engagement. However, there remains a lack of a comprehensive model to guide such developments. This study aims to propose the Conversational AI-powered Virtual Reality Development Model for Tourism Promotion in Thailand, providing an integrated and context-specific framework suitable for practical implementation. A Design and Development Research (DDR) methodology (Type II) was employed in three stages: 1) synthesizing essential components through a scoping review, 2) constructing and validating the model via expert panels using the Content Validity Index (CVI) analysis, and 3) assessing suitability and acceptance through expert evaluation and stakeholder surveys. The model developed in this study, referred to as the 4Ds Model, contributes new knowledge by integrating conversational AI and virtual reality within a four-phase structure—Discover, Design, Develop, and Deploy—supported by five enabling capitals: human, cultural, technological, informational, and financial. The Deploy phase modifies the AISAS communication framework into AICAS (Attention, Interest, Chat, Action, Share) to illustrate the function of conversational AI in improving user interaction and engagement within the context of tourism in Thailand. Results indicated high expert ratings of suitability and strong stakeholder intention to adopt. Multiple regression analysis revealed that technological self-efficacy, perceived interactivity, and perceived tourism benefits were significant predictors, explaining 73.3% of the variance in behavioral intention. The findings demonstrate both the theoretical advancement in AI–VR integration and the practical readiness of the 4Ds Model as a culturally aligned roadmap for digital tourism transformation in Thailand.

Author 1: Jenasama Srihirun

Author 2: Kridsanapong Lertbumroongchai

Author 3: Vitsanu Nittayathammakul

Author 4: Pimon Kaewdang

Keywords: Virtual reality; conversational AI; Model Development; digital tourism; technology adoption

Paper 74: Roadmap for Emerging Cyberbullying Mitigation: Integrating AI-Based Solutions, Ethics, and Policy

Abstract: Cyberbullying is one of these challenges that are most found among the younger users of social media which affects the mental health. Artificial Intelligence (AI) is rapidly developing and has enormous potential to mitigate cyberbullying. Therefore, this chapter will talk about the role AI has started playing in strengthening the efforts to combat cyberbullying. Cyberbullying includes all forms of deliberate aggressive behaviour that aims to inflict social, psychological or physical pain in a digital space and AI detection technologies have a lot of potential to detect, predict and prevent cyberbullying in real time. Other critical components of the chapter are how the advances in Natural Language Processing (NLP) technologies, machine learning, images and videos, behavioural analytics make AI an emerging innovation to prevent cyberbullying and provide better services in a timely manner. There are positive trends that make it clear how Safer AI can help in improving the safety of future digital environments. More advanced NLP models will be able to identify the nuances of cyberbullying involving indirect attacks and sarcasm. The chapter will also discuss the hazards associated with AI-based solutions, such as privacy, the zero-sum game of AI morality against AI effectiveness, and the importance of explaining and assigning responsibility for every AI decision. It shows how AI is changing our approach to online safety and helps us identify cyberbullying in a variety of media, including text, video and images. This article gives an overview of roadmap for cyberbullying mitigation with the assistance of AI and ethical practices.

Author 1: Atif Mahmood

Author 2: Shaik Shabana Anjum

Author 3: Umm E Mariya Shah

Author 4: Pavani Cherukuru

Author 5: Javid Iqbal

Author 6: Sarah Bukhari

Keywords: Cyberbullying; human computer interaction; artificial intelligence; natural language processing; machine learning; content moderation; predictive analytics; online safety; youth protection; mental health; mental illness; cybercrime

Paper 75: Intelligent Visualization and Knowledge Graph Analysis for Trend Detection

Abstract: This research employs scientometric examination and visual analytics techniques anchored in the Web of Science (WoS) repository to methodically delineate predominant research themes, foundational academic works, and emerging scholarly directions within industry-education integration studies. The investigation seeks to elucidate the discipline's epistemological framework and longitudinal transformation patterns while offering innovative analytical lenses and methodological paradigms to advance theoretical conceptualization and operational innovation in industry-education convergence initiatives. This investigation employs scientometric techniques to systematically map and examine 500 scholarly works on industry-education integration from the Web of Science (WoS) database (2010–2023) using VOSviewer. Through co-occurrence mapping, thematic clustering, and temporal trend analysis, the study identifies dominant research foci, influential contributors, and collaborative networks. This quantitative approach is further supplemented by case study investigations to delineate operational strategies and innovative frameworks for industry-academia synergy. Analysis reveals that research concentration spans five domains: higher education reform, Industry 4.0 alignment, engineering pedagogy enhancement, innovation ecosystems, and sustainability integration. Temporal evolution tracking demonstrates a paradigm shift from foundational theoretical debates to applied technological and implementation studies in recent cycles. Cluster analytics highlight the interdisciplinary nature of industry-education convergence, emphasizing tripartite collaboration among academic institutions, corporate entities, and governmental bodies as pivotal to systemic advancement. By synthesizing research trajectories and thematic priorities, this work establishes a structured knowledge foundation for both theoretical refinement and practical implementation in industry-education integration.

Author 1: Sunan Lv

Keywords: Reviewer; industry-education integration; hotspot visualization and analysis

Paper 76: An Efficient and Scalable Reinforcement Learning-Driven Intelligent Resource Management and Secure Framework for LoRaWAN

Abstract: This study proposes a Q-learning-based adaptive duty cycle scheduling algorithm for LoRaWAN in a smart city eco-system to enhance the energy efficiency, reduce transmission delay, and handle dynamic traffic conditions. Additionally, it also incorporates an intelligent and efficient channel utilization scheme for LoRaWAN-enabled IoT networks and also integrates a lightweight security strategy at the edge (gateways), making it suitable for low-power, low-computation LoRaWAN environments. In this adaptive and intelligent LoRaWAN framework Q-learning agent dynamically selects various transmission actions based on the contextual states, including buffer size, energy levels, and channel conditions, which optimizes energy efficiency and also enhances the reliability of data transmission in LoRaWAN. The light-weight intrusion detection mechanism also filters suspicious packets using trust scores and payload analysis to ensure secure data delivery and adaptive, scalable, and proactive protection against several prevalent threats in LoRaWAN-driven IoT. It also incorporates a channel-aware scheduling to avoid congestion and improve overall transmission performance. Experimental outcome further confirms improvement over throughput, delay, bandwidth utilization, energy conservation, and resilience against malicious or faulty transmissions, demonstrating the framework’s ability to optimize the resource allocation performance while balancing the above metrics adaptively.

Author 1: Shaista Tarannum

Author 2: Usha S. M

Keywords: LoRa; LoRaWAN; Q-learning; adaptive duty cycle; channel scheduling; energy efficiency; intrusion detection; trust score; resource management; IoT security

Paper 77: Real-Time Multi-Scale Object Detection in Surveillance Using Hybrid Transformer Architecture

Abstract: Real-time surveillance systems require accurate and efficient object detection to ensure safety and situational awareness. Existing methods, such as YOLOv5 and Vision Transformer-based detectors, often struggle to reliably identify small, distant, or occluded objects while maintaining real-time inference, limiting their applicability in complex surveillance environments. To address these challenges, this study proposes PRISM, a hybrid Transformer–YOLOv8 framework that integrates fast local feature extraction with global contextual refinement. The method introduces two novel components: i) a Context-Aware Feed Forward Network (CA-FFN) within the Vision Transformer (ViT), which dynamically weights channel features to reduce redundancy and enhance global context modeling, and ii) Cross-Scale Attention Skip Connections (CSASC) for selective fusion of multi-scale YOLOv8 and ViT features, improving detection of small or occluded objects. The model is implemented in PyTorch and trained on a comprehensive surveillance dataset consisting of pedestrians, vehicles, bicycles, bags, and miscellaneous objects. Experimental evaluation demonstrates that PRISM achieves 96% accuracy, a significant improvement of ~4–5% over baseline methods, with robust performance across all object categories. Key performance indicators verify the reliability of the model to real-time usage, and the lightweight design makes it edge deployable. These findings imply that PRISM can be used to provide a speed-accuracy balance in a complex and dynamic setting, which is more efficient than the current methods. The study also notes the partial extensions, such as the incorporation of multi-sensors and continuous video streams to do time modeling as an extension, which will offer a good base to the next-generation intelligent surveillance systems.

Author 1: Roshan D Suvaris

Author 2: Rahul Suryodai

Author 3: S. Narayanasamy

Author 4: Aanandha Saravanan

Author 5: Raman Kumar

Author 6: P N V Syamala Rao M

Author 7: Elangovan Muniyandy

Keywords: Real-time object detection; hybrid transformer–YOLOv8; Context-Aware Feed Forward Network (CA-FFN); Cross-Scale Attention Skip Connections (CSASC); surveillance video analytics; multi-scale feature fusion

Paper 78: Aerial Draft Surveyor (ADS)

Abstract: Draft surveying is an essential procedure in determining the displacement and loaded cargo weight of bulk carriers. Currently, the most acceptable method is through manual visual observation by trained draft surveyors. However, this process is subjective, error-prone, and unsafe under poor visibility or during rough sea conditions. This study presents an automated computer vision-powered UAV draft surveying system integrating TensorRT Optimized YOLO11n object detection and YOLO11n-seg image segmentation models deployed on an NVIDIA Jetson Orin Nano. The system performs real-time draft estimation by detecting draft marks, segmenting the waterline, and computing draft values using convergence and line-fitting algorithms. Comparative evaluation with licensed human surveyors on 40 paired readings yielded an MAE of 0.1068 m, RMSE of 0.2740 m, and an R² of 0.948, demonstrating human-comparable accuracy. Agreement analysis indicates high reliability (two-way random effects ICC(2,1) = 0.974) and a small mean bias (system − manual = +0.0628 m, 95% limits of agreement: −0.467 m to +0.592 m). Moreover, a paired t-test (t = 1.469, df = 39) found no statistically significant difference between methods (p ≈ 0.150). The results validate that the proposed UAV-driven computer vision system can perform reliable, real-time draft surveying with accuracy comparable to human experts.

Author 1: John Matthew H. Escarro

Author 2: Fharjan M. Taguinopon

Author 3: Gyrielle Kysha M. Demegillo

Author 4: Dan Kevin T. Amper

Author 5: Rosanna C. Ucat

Author 6: Mark John S. Pag-Alaman

Keywords: Draft survey; UAV; machine learning; computer vision

Paper 79: Privacy-Aware Federated Graph Neural Networks for Adaptive and Explainable Cancer Drug Personalization

Abstract: Personalized cancer treatment remains challenging due to the complexity of genomic data and variability in drug responses. Previous federated learning (FL) approaches handled distributed patient data to preserve privacy but treated genomic and pharmacological features as flat, tabular inputs, limiting the ability to capture gene–drug interactions. In this study, we propose a Graph Neural Network (GNN)-based framework, FedGraphOnco, which models patient-specific gene–drug interactions as structured graphs, enabling the network to learn complex relational patterns that are difficult or impractical for FL-only models. Attention mechanisms and SHapley Additive exPlanations (SHAP) are incorporated to provide interpretable insights into important genes, pathways, and drug interactions, increasing clinical trust. Using the GDSC dataset with gene expression, mutation status, copy number variation, and IC50 drug responses, the model demonstrates high predictive accuracy (Pearson correlation = 0.85, RMSE = 2.6, MAE = 1.9, dosage deviation = 2.8%), robustness to noise and non-IID data, and adaptive, personalized dosage recommendations. The approach highlights the advantages of combining privacy-preserving FL, GNNs, multi-omics data integration, explainability, and adaptive dosing, offering a scalable and interpretable solution for precision oncology.

Author 1: Tripti Sharma

Author 2: Lakshmi K

Author 3: M. Misba

Author 4: Jasgurpreet Singh Chohan

Author 5: R. Aroul Canessane

Author 6: Komatigunta Nagaraju

Author 7: Adlin Sheeba

Keywords: Graph Neural Networks; cancer drug dosage; privacy preservation; genomic profiling; precision oncology

Paper 80: Hybrid Vision Transformer and MLP-Mixer for Epileptic Seizure Detection in Intracranial EEG

Abstract: Accurate and timely seizure detection is essential for effective epilepsy management, and automated systems can play a valuable role in supporting clinical practice. In this study, we introduce a hybrid approach that uses time-frequency representations of Intracranial electroencephalography (iEEG) signals filtered at High-Frequency Oscillations (HFOs) bands as input to different convolutional neural network (CNN) backbones for feature extraction, followed by classification with either a Vision Transformer (ViT) or MLP-Mixer. This work establishes a systematic, comparative framework for benchmarking hybrid CNN-ViT against CNN-MLP-Mixer, providing a critical new reference for automated epileptic seizure detection within HFOs filtered iEEG signals. Extensive evaluation demonstrates that the ViT consistently achieves superior performance, with an EfficientNetB0-ViT model attaining remarkable accuracy (97.85%) and specificity (98.92%). Crucially, the MLP-Mixer emerges as a highly competitive alternative, exhibiting strong recall capabilities that make it suitable for applications where missing a seizure is not an option. Overall, our findings suggest that self-attention mechanisms in ViTs provide a distinct advantage for capturing complex seizure dynamics, yet MLP-based models present a powerful, efficient option.

Author 1: Thouraya Guesmi

Author 2: Abir Hadriche

Author 3: Nawel Jmail

Keywords: Vision transformer; MLP-Mixer; iEEG; HFOs; ResNet; GoogleNet; EfficientNetB0

Paper 81: Truth Under Pressure: A Deep Learning-Based Lie Detection System for Online Lending Using Voice Stress and Response Latency

Abstract: The rapid increase in defaults in the online lending industry highlights significant flaws in current debtor verification, which largely relies on static, preparable interviews, leading to high non-performing loans. Existing research is fragmented: while Large Language Models (LLMs) show promise in question generation, their application is confined to non-financial domains like education, and lie detection studies often analyze modalities in isolation. This study addresses this critical gap by proposing the first integrated AI-driven system for this context. We solve the problem in two parts: 1) A Llama 3 LLM is fine-tuned to generate dynamic, biodata-tailored questions, preventing the rehearsed answers that plague static interviews. 2) A novel multimodal deep learning model is developed to analyze the response, uniquely fusing vocal acoustic features and response latency—two key deception indicators that prior work has failed to combine. The Llama 3 model produced a low perplexity score (2-3), and the lie detection model achieved 70% testing accuracy with a 70.9% F1-Score. Despite signs of overfitting, this framework provides a novel, intelligent decision-support tool to reduce fraud and manage default risks more effectively.

Author 1: Ahmad Ihsan Farhani

Author 2: Alhadi Bustamam

Author 3: Rinaldi Anwar

Author 4: Titin Siswantining

Keywords: Online lending; lie detection; large language model; deep learning; voice acoustics; response latency

Paper 82: Classification of Mangrove Ecosystem Health Using Sentinel-2 Images with Genetic Algorithm Optimization in Machine Learning Algorithms

Abstract: Mangrove ecosystems play an important role in maintaining coastal ecological balance, including as carbon sinks and natural protection from abrasion, but mangrove areas in Mempawah Regency have experienced significant degradation due to anthropogenic pressures. Therefore, this study aims to classify the health condition of mangroves using multi-temporal Sentinel-2 imagery with a hybrid machine learning (ML) approach and Genetic Algorithm (GA) optimization. We implemented GA optimization comparatively on four main ML models—Multilayer Perceptron (MLP), Decision Tree (DT), XGBoost, and Naïve Bayes (NB)—to adjust hyperparameters to improve accuracy and reduce overfitting. The results prove that GA optimization effectively improves classification performance, with the MLP-GA model providing the highest accuracy with an increase of up to 3.8% compared to the non-optimized baseline model, achieving a best performance value of ROC AUC 0.9730 and reducing computation time by up to 60%. These findings indicate that the GA-MLP framework is highly reliable and efficient, providing a precise tool for strategic decision-making in the management of healthy mangrove ecosystems.

Author 1: Putri Yuli Utami

Author 2: Murni Ramadhani

Author 3: Rudi Alfian

Author 4: Barry Ceasar Octariadi

Author 5: Dimas Kurniawan

Keywords: Classification; genetic algorithm; machine learning; mangrove ecosystem; Sentinel-2

Paper 83: User Identity Confirmation Property Management System Based on State Secret Algorithm and Blockchain Technology

Abstract: The existing user identity confirmation methods in property management systems are vulnerable to attacks and forgery, posing serious threats to system security and reliability. To address these issues, this study proposes a novel user identity confirmation method that combines the state secret SM9 algorithm with blockchain technology. The system utilizes blockchain for managing and verifying user identity information, while employing the SM9 algorithm for double encryption of user data. This approach ensures robust protection against identity theft and fraud, enhancing security and privacy. The proposed method was tested experimentally, and the results show that the model achieves an average communication connection and verification initiation time of approximately 11.07 ms, with a key negotiation success rate of 88.73%. Moreover, the model achieved a user identity confirmation accuracy of 90.41%, which is significantly higher than traditional methods. These findings highlight that the integration of the SM9 algorithm and blockchain technology offers high accuracy, low latency, and improved scalability, making it an ideal solution for enhancing the security and efficiency of property management systems.

Author 1: Xiao Tian

Author 2: Xing Chen

Keywords: User identification; state secret algorithm; blockchain technology; property management system; security

Paper 84: Coordination, Communication and Robustness in Multi-Agents: An Industrial Network Scenario Using Trust Region Policy Optimization

Abstract: Numerous practical uses necessitate multi-agent systems, including managing traffic, assigning tasks, regulating ant colonies, and operating self-driving cars, and drones. These systems involve multiple agents working together, communicating and engaging with their surroundings to achieve the highest possible total numerical reward. Deep Reinforcement Learning (DRL) approaches are used to address these multi-agent applications. In many circumstances, the use of agents raise challenges to safety and robustness. To address these issues, we develop a DRL based system in which multiple agents in an industrial network scenario interact with the real-world environment and act collaboratively and cooperatively. In proposed model, several agents collaborate with one another to complete tasks and maintain a safe state. To take actions cooperatively and collaboratively of agents in accordance with the safety robustness of policies, we apply DRL algorithms such as proximal policy optimization (PPO) and Trust Region Policy Optimization (TRPO) algorithms and DRL approaches. We apply Curriculum Learning (CL) for their better performance and training. In this study, a reward structure is also proposed which help agents to maintain their safe state. Mean reward, policy loss, value loss, value estimate and safety robustness are analyzed as performance matrix in this study. The results shows that the policy adopted in the proposed model perform comparably better than the other policies.

Author 1: Munam Ali Shah

Keywords: Safety robustness; reinforcement learning; multi-agents; safe state; collaboration

Paper 85: Spatiotemporal Graph Networks for Relational Reasoning in Campus Infrastructure Management

Abstract: The efficient management of campus infrastructure presents a complex spatiotemporal forecasting challenge characterized by dynamic interdependencies between physical assets. Traditional models fail to capture these intricate relationships as they treat buildings as independent entities or rely on static correlation structures. This paper introduces a novel Spatiotemporal Graph Neural Network (ST-GNN) framework that reframes infrastructure forecasting as a relational reasoning task, enabling dynamic inference of campus wide interdependencies. Our approach integrates Graph Attention Networks (GAT) to learn time-varying spatial dependencies and Gated Temporal Convolutional Networks (TCNs) to capture multi-scale temporal patterns. A key innovation is our context-sensitive graph construction method that incorporates physical proximity, functional similarity, and human mobility data to create a holistic representation of campus dynamics. Evaluated on a real-world multimodal dataset comprising 24 months of energy and occupancy data from 50 campus buildings, the proposed model demonstrates superior performance, achieving a 16.3% reduction in mean absolute error compared to the strongest baseline. Comprehensive ablation studies confirm the critical contribution of each architectural component, while qualitative analysis reveals the model’s capacity to provide interpretable insights into campus operational patterns. This work provides a powerful framework for intelligent campus management, enabling precise resource allocation, energy optimization, and sustainable operational planning through advanced relational reasoning capabilities.

Author 1: Sanjay Agal

Author 2: Krishna Raulji

Author 3: Nikunj Bhavsar

Author 4: Pooja Bhatt

Keywords: Spatiotemporal Graph Neural Networks; relational reasoning; smart campus management; infrastructure utilization forecasting; graph attention networks; temporal convolutional networks; dynamic graph construction; energy optimization; predictive analytics

Paper 86: A Voting-Based Ensemble Method for Deep Learning Performance Enhancement

Abstract: Overfitting and limited generalization remain significant challenges for deep learning models, often leading to suboptimal performance on unseen data. To address this, The Divided Ensemble Voting (DEV) method was introduced, a novel approach that strategically partitions a dataset into distinct sub-sets to train an independent model on each partition. This division encourages each model to specialize in unique features and patterns, thereby increasing ensemble diversity. Predictions from all models are aggregated through a majority voting mechanism to determine the final output, which mitigates overfitting and improves generalization. The proposed method was rigorously evaluated on four binary image classification tasks: Deepfake & Real, Waste Classification, Concrete & Pavement Crack, and Non & Biodegradable Material. Experimental results demonstrate that DEV consistently surpasses the performance of conventional singular models. Accuracy rates improved from 85.55% to 93.1%, 85.12% to 89.6%, 95.42% to 99.0%, and 89.00% to 93.0%, respectively, across the datasets. These findings underscore the efficacy of strategic data partitioning and ensemble consensus in advancing deep learning performance.

Author 1: Mohammed Abdel Razek

Author 2: Rania Salah El-Sayed

Author 3: Arwa Mashat

Author 4: Shereen A. El-aal

Keywords: Divided Ensemble Voting (DEV); deep learning (DL); CNN; binary classification; performance metrics

Paper 87: A Spatiotemporal Forex Trading System Based on a Hybrid Model GAT-LSTM: Forecasting Forex Price Directions

Abstract: Due to the high volatility and complex interdependencies within financial markets, predicting Forex prices becomes a difficult challenge for investors. Furthermore, the traditional trading models struggle to capture those relationships. To address this issue, we introduced a spatiotemporal Forex trading system, GAT-LSTM-based; it is a hybrid approach that combines Graph Attention Network (GAT) with a Long Short-Term Memory (LSTM) network. The GAT component helps to capture spatial dependencies between currencies by constructing a directed graph containing 28 currency pairs alongside commodity stock and US stocks. The strength of the GAT component lies in its ability to dynamically adjust and recalculate the weights of edges over time, which helps our proposed system to adapt to macroeconomic changes, news events, and financial factors that can impact the Forex market status. The LSTM component deals with the nature of time series datasets. It learns temporal interdependencies, allowing our system to detect repeated long-term patterns over time. Experimental results proved that the suggested hybrid model, GAT-LSTM, surpasses both LSTM and GAT separately. By combining both elements and leveraging simultaneously the strength of dynamically modelling spatial dependencies, and the strength of learning long-term temporal patterns, our suggested system became more accurate in forecasting Forex price directions, showing promising results and high accuracy during the validation phase.

Author 1: Nabil MABROUK

Author 2: Marouane CHIHAB

Author 3: Younes CHIHAB

Keywords: Forex trading; hybrid deep learning model; Graph Attention Network (GAT); Long Short-Term Memory (LSTM); spatiotemporal forecasting

Paper 88: Experimental Validation of an Adaptive Controller for a Mecanum-Wheel Robot with Unknown Center-of-Gravity Offset and Slope Inclination

Abstract: High-precision path tracking for a Four-Mecanum-Wheel Mobile Robot (FMWMR) is challenged by real-world factors such as payload-induced shifts in the center-of-gravity (CoG) and operation on inclined surfaces. These uncertainties introduce complex, coupled dynamic forces that degrade the performance of conventional controllers. This study addresses this problem with a Model Reference Adaptive Controller (MRAC), which learns and compensates for these unpredictable dynamic effects in real time. To ensure effective operation on physical hardware, the controller incorporates practical solutions for motor friction and control signal stability. The proposed approach is validated through implementation of the MRAC on a Rosmaster X3 robot. A performance comparison is made against a well-tuned Proportional-Integral-Derivative (PID) controller is conducted across twelve distinct scenarios. The results show that the adaptive controller reduced the position Root Mean Square Error (RMSE) by an average of 52.7% and the Integral of Time-weighted Absolute Error (ITAE) by 61.5%. This work validates the MRAC as a powerful and robust solution for robots operating in unpredictable environments.

Author 1: Chawannat Chaichumporn

Author 2: Supaluk Prapan

Author 3: Nghia Thi Mai

Author 4: Md Abdus Samad Kamal

Author 5: Iwanori Murakami

Author 6: Kou Yamada

Keywords: Model reference adaptive control; Mecanum Wheel Robot; center of gravity offset

Paper 89: Enhancing Out-of-Distribution Detection for Retail Time-Series Data Using Entropic Methods

Abstract: Machine learning models are typically developed under the “closed-world” assumption, where training and testing data originate from a consistent distribution. However, in real-world scenarios, especially in the retail domain, this assumption can become problematic due to the frequent introduction of new products, seasonal promotions, and irregular sales events. When models encounter out-of-distribution data inputs, predictions can become overly confident or entirely incorrect. While existing out-of-distribution detection methods primarily focus on image-based datasets, challenges associated with numerical, high-dimensional, and heterogeneous retail time-series data remain largely unexplored. To address this gap, this study proposes an enhanced Entropic Out-of-Distribution Detection framework tailored specifically for dynamic retail environments. By trans-forming time-series sales data into spectrogram representations and leveraging the IsoMax+ loss function, our approach im-proves uncertainty calibration and robustness without requiring labeled out-of-distribution data or additional post-hoc calibration techniques. Experimental results, conducted on a large-scale retail dataset from Vietnam, demonstrate that the proposed Entropic Out-of-distribution detection framework significantly outperforms traditional out-of-distribution detection methods in terms of detection accuracy and inference efficiency, providing a scalable and practical solution for real-time retail applications. Our approach achieves strong performance with an F1-score of 88% and an AUC of 91%, highlighting its promising applicability across diverse business scenarios.

Author 1: Nga Nguyen Thi

Author 2: Tuan Vu Minh

Author 3: Khanh Nguyen-Trong

Keywords: Out-of-Distribution Detection; entropic learning; IsoMax+ loss; time-series classification; retail forecasting; deep learning; spectrogram transformation

Paper 90: A Review of Artificial Intelligence in Inventory Management: Methods, Applications and Directions

Abstract: Effective inventory management is fundamental to supply chain resilience and efficiency. Artificial intelligence (AI) has emerged as a transformative solution that enables more dynamic and data-driven inventory strategies. To map the latest advancements in this rapidly evolving field, this study presents a systematic literature review (SLR) of AI techniques in inventory management. The review was conducted following the PRISMA 2020 guidelines, through which 87 high-quality articles published between 2021 and 2025 were systematically analyzed. Our review identifies machine learning (ML), deep learning (DL), reinforcement learning (RL), and hybrid methods as the predominant AI technologies. These techniques primarily address three foundational tasks. In demand forecasting, they improve prediction accuracy and mitigate stockout and overstock risks. For inventory control, they balance costs with service levels and optimize replenishment strategies. In inventory classification, they facilitate targeted resource allocation. Despite these advancements, AI research confronts significant challenges, particularly in data dependency, model interpretability, and implementation overhead. To address these gaps, we suggest future research focused on data-efficient learning, explainable AI, and lightweight, integrated frameworks to lower adoption barriers. This review provides a timely and holistic overview of the current research landscape, which serves as a reference for academics to identify research directions.

Author 1: Jinjin Li

Author 2: Huijun Huang

Author 3: Yuping Gong

Author 4: Lei Wang

Author 5: Xiangui Yin

Author 6: Yichang Liu

Keywords: Inventory management; artificial intelligence; demand forecasting; inventory control; inventory classification; machine learning

Paper 91: Glioma Classification Using Harris Hawks-Driven Optimized Gradient Boosting Classifier Along with SHAP-Based Interpretability

Abstract: Gliomas are considered one of the most lethal and aggressive types of brain cancer, responsible for countless deaths worldwide. This study seeks to improve glioma classification using cutting-edge machine learning (ML) techniques to differentiate between glioma subtypes based on clinical and genomic data. The goal is to identify important biomarkers and features influencing glioma classification, with an emphasis on improving feature selection and model interpretability. For glioma classification, the Gradient Boosting Classifier (GBC) was employed. The Harris Hawks Optimization (HHO) algorithm was used for feature selection and hyperparameter fine-tuning to enhance the model’s performance. Additionally, SHapley Additive exPlanations (SHAP) were applied to improve model interpretability and to better understand feature contributions.The Gradient Boosting (GB) method yielded the best performance among the selected models, achieving an accuracy of 88.40%, precision of 87.3%, recall of 88.48%, and an F1 score of 88.29%, with feature selection and hyperparameter tuning using the Harris Hawks Optimization. These results highlight the significance of hyperparameter tuning and feature selection in enhancing classification performance. Key features such as IDH1, Age at Diagnosis, and EGFR were identified as the most influential in distinguishing glioma subtypes. SHAP analysis further confirmed the importance of these features in the model.This study shows that the Gradient Boosting Classifier (GBC), optimized with Harris Hawks Optimization (HHO), significantly improves glioma classification, achieving a high F1 score. Key features like IDH1, Age at Diagnosis, and EGFR were identified, showcasing its potential for enhanced glioma diagnosis.

Author 1: SM Naim

Author 2: Jun-Jiat Tiang

Author 3: Abdullah-Al Nahid

Keywords: Glioma; gradient boosting; Harris Hawks Optimization (HHO); SHAP; feature selection; interpretability; TCGA; IDH1; EGFR

Paper 92: Federated Performance-based Averaging (FedPA): A Robust and Selective Learning Framework for Chest X-Ray Classification in Heterogeneous Data Environments

Abstract: Chest X-ray imaging remains a cornerstone in the diagnosis of thoracic conditions such as COVID-19, pneumonia, and lung opacity. Despite advancements in deep learning, the development of robust and generalizable models is limited by data privacy constraints, as patient data cannot be centralized across institutions. Federated Learning (FL) has emerged as a promising solution by enabling collaborative model training without sharing raw data. However, standard FL algorithms like FedAvg, FedProx, and FedSGD aggregate all client updates without considering their individual quality, making them vulnerable to performance degradation in the presence of data heterogeneity, label noise, or underperforming clients. To address these challenges, this study proposes Federated Performance-Based Averaging (FedPA), a novel selective aggregation strategy that incorporates only those client models that meet a pre-defined performance threshold during training. By leveraging an accuracy-based filtering mechanism, FedPA ensures that only sufficiently trained and reliable local models contribute to global updates. The method was evaluated on a multi-class, non-IID chest X-ray dataset containing four classes: Normal, COVID-19, Pneumonia, and Lung Opacity. Using DenseNet as the backbone model, experiments were conducted across four federated clients, each biased toward a specific class to simulate real-world data distributions. Results demonstrate that FedPA significantly outperforms baseline federated algorithms across key metrics, achieving a global accuracy of 91.82%, F1-score of 92.48%, and recall of 92.08%. The method also achieved faster convergence, higher stability, and reduced round-to-round accuracy fluctuations. System-level evaluations further show that FedPA offers competitive efficiency in terms of inference time, throughput, CPU usage, and memory footprint, making it suitable for deployment in resource-constrained clinical environments. Overall, FedPA offers a practical and effective advancement in federated learning for medical imaging. By filtering unreliable client contributions, it preserves model quality and privacy, presenting a viable path for clinical deployment in scenarios where data centralization is infeasible due to ethical, legal, or logistical constraints.

Author 1: Atif Mahmood

Author 2: Tashin Khan Sadique

Author 3: Saaidal Razalli Azzuhri

Author 4: Roziana Ramli

Author 5: Leila Ismail

Keywords: Public health; industrial growth; federated learning; FedAvg; FedPA; FedSGD

Paper 93: Privacy-Preserving Education Data Sharing Scheme Based on Consortium Blockchain

Abstract: With the growing emphasis on lifelong education and the rapid expansion of open education platforms, the secure and efficient management and sharing of lifelong learning data have become critical challenges. To address these issues, this paper proposes a Privacy-Preserving Educational Data Sharing (PPEDS) scheme based on blockchain technology. The PPEDS scheme employs attribute-based encryption with hidden attributes to achieve privacy-preserving and fine-grained access control. In addition, it incorporates multi-keyword searchable encryption to enable efficient encrypted data retrieval and combines private and consortium blockchains to ensure data authenticity and integrity across multiple educational institutions. The security analysis demonstrates that the scheme resists potential attacks and ensures confidentiality, access control, and search privacy under a semi-trusted model. Furthermore, performance evaluations conducted on real-world educational datasets show that the proposed scheme achieves efficient encryption, search, and decryption operations, with low computational overhead even in large-scale deployments. Overall, the PPEDS scheme provides a secure, scalable, and practical solution for privacy-preserving data sharing in lifelong education systems.

Author 1: Jiaqi Guo

Author 2: Zhuoran Wang

Author 3: Ningning Liu

Keywords: Consortium blockchain; access control; secure search; life-long education; data sharing

Paper 94: RT-DETR Edge Deployment: Real-Time Detection Transformer for Distracted Driving Detection

Abstract: Distracted driving is one of the primary contributors to road accidents worldwide, highlighting the urgent need for reliable in-cabin driver monitoring systems. Existing approaches often face trade-offs: CNN-based classifiers achieve high recognition accuracy but lack spatial localization, while lightweight real-time detectors sacrifice contextual reasoning for efficiency. To bridge this gap, we propose a customized fine-tuned transformer-based object detection framework, RT-DETR-L, specifically adapted for distracted driving detection. In contrast to prior applications of RT-DETR, our adaptation integrates distraction-specific data augmentation, loss-balancing strategies, and deployment-oriented optimizations, enabling precise classification and spatial localization of distractions such as texting, drinking, yawning, and eye closure. Trained and validated on a large-scale annotated in-cabin dataset, RT-DETR-L achieves state-of-the-art performance with a mAP50 of 0.995 and mAP50–95 of 0.774. In addition the proposed model demonstrates the deployment feasibility on resource-constrained embedded platforms (ARM-based edge AI devices), where the model sustains real-time performance at 17.5 FPS with minimal latency. These results establish RT-DETR-L as a hybrid solution combining the semantic depth of transformers with the efficiency required for Advanced Driver Assistance Systems (ADAS). By addressing both accuracy and deployability, this study makes concrete contributions toward advancing robust, real-time driver monitoring for enhanced road safety.

Author 1: Fares Hamad Aljahani

Keywords: RT-DETR; real-time inference; autonomous vehicles

Paper 95: Breast Cancer Classification Using Ensemble Voting: A Feature Selection Approach

Abstract: Breast cancer is one of the most common and deadly diseases affecting women around the worldwide. It is specially affecting in regions where has limited access to advanced diagnostic tools. Recent studies have shown that blood-based biomarkers can give a cost-effective alternative for early detection. This paper represents a machine learning-based approach for classifying breast cancer using clinical and biomedicial data. We have used the Breast Cancer Coimbra dataset for our study. We employed four filter-based feature selection methods—Mutual Information, Chi-Square, ANOVA F-test, and Pearson Correlation Coefficient—to identify the most relevant features for classification. We have applied two classifiers (AdaBoost and Ensemble Voting Classifier) to enhance predictive accuracy. The ensemble model achieved an accuracy of 82.86%. Key features such as glucose, HOMA, insulin, resistin, and age consistently contributed across all selected methods.It highlights that a few of the features has a great significance in breast cancer prediction. This study also try to investigate the reasons behind the missclassification cases. Our results show that using statistical feature selection with ensemble learning reasonable helps to boost the accuracy of breast cancer prediction. This approach helps the model focus on the most important features.

Author 1: Antu Kumar Guha

Author 2: Jun-Jiat Tiang

Author 3: Abdullah-Al Nahid

Keywords: Breast cancer; machine learning; feature selection; ensemble learning; AdaBoost; biomedical data classification

Paper 96: A Two-Stage Framework for Abnormalities Detection in WCE Images by Combining Semantic Segmentation and Deformable Agent-Based Classification

Abstract: Wireless capsule endoscopy (WCE) has revolutionized gastrointestinal (GI) diagnostics by offering a patient-friendly imaging and diagnostic tool compared to traditional endoscopic techniques. However, the manual assessment of these images is a time-consuming task and is prone to inaccuracies, which necessitates the implementation of automated approaches. In this paper, we introduce a two-stage deep learning framework to identify the most common GI abnormalities in WCE images. The first stage in the proposed method is to segment suspicious regions from the WCE images, which act as potential markers for GI abnormalities. In the second stage we perform frame-level classification to identify and categorize different pathologies in the GI tract. Extensive experiments conducted on four image datasets demonstrate that our approach achieves the highest values in terms of accuracy, precision, recall and specificity in comparison with four common deep learning methods : Resnet50, VGG16, Vit-S16 and InceptionV3.

Author 1: Brahim Alibouch

Author 2: Yasmina El Khalfaoui

Keywords: Wireless capsule endoscopy; deep learning; classification; gastrointestinal abnormalities

Paper 97: Predicting Employee Attrition in the Saudi Private Sector Using Machine Learning

Abstract: Employee attrition represents a prominent issue facing organizations, as human capital represents one of the most valuable resources. Attrition refers to the voluntary or involuntary reduction in the number of employees, which can negatively affect profitability, reputation, and overall organizational performance. Therefore, a comprehensive understanding of this phenomenon, its causal factors, and the development of retention strategies is crucial for mitigating employee turnover. The purpose of this work is to predict employee attrition in the Saudi private sector and identify the key factors contributing to employee turnover using machine learning approaches. in addition, the research structurally evaluates the performance of multiple Machine Learning (ML) algorithms within the proposed framework to determine the most effective predictive model for employee attrition. This study utilized a training dataset obtained from an online survey targeting employees in the Saudi private sector in order to investigate employee attrition and identify its most prominent causes within this context. Thus, various Machine Learning (ML) algorithms, including Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Bagging ensemble, and Voting Classifier (VC) were evaluated. The results demonstrate that the Voting Classifier (VC) yielded the highest accuracy at 90%. Moreover, the analysis identified job opportunities and job titles as some of the most influential factors driving employee turnover.

Author 1: Haya Alqahtani

Author 2: Hana Almagrabi

Author 3: Amal Alharbi

Keywords: Employee attrition; attrition prediction; predictive models; machine learning; voting classifier; ensemble methods; Saudi private sector; employee turnover; employee retention; feature importance

Paper 98: Optimizing Asset Transfer Process in ERP Using Business Process Management Technique

Abstract: Enterprise Resource Planning (ERP) systems are critical for managing enterprise-wide business processes, including asset management. Yet, many ERP platforms lack efficient mechanisms for bulk asset transfers, leading to high manual effort, increased costs, and data inconsistencies. This study applies Business Process Reengineering (BPR) techniques as the methodology to optimize ERP asset management, focusing on workflow optimization and automation, contributing both practical and methodological insights. A mixed-method approach was adopted, analyzing a financial organization with 256 branches and over 450 Oracle ERP users. Data from 51 representative branches identified inefficiencies such as manual transfer delays, approval bottlenecks, and synchronization issues. The proposed solution introduces automated bulk asset transfers, optimized approval workflows, and real-time data synchronization, along with new metrics for evaluating efficiency, compliance, risk, and asset utilization. Compared to the As-Is system, the reengineered framework achieved a 100% reduction in operational costs per user ($7,500 annual saving), an 80% reduction in compliance incidents, a 67% reduction in asset transaction errors, and a 20% improvement in asset utilization. These results demonstrate a scalable, adaptable, and effective framework that enhances ERP operational efficiency, strengthens data integrity, and advances both academic understanding and industrial practice of asset management process reengineering.

Author 1: Ravindu Yasarathne

Author 2: Naduni Ranatunga

Author 3: Vikasitha Herath

Author 4: Lakshan Chalinda

Author 5: Chathurangika Kahandawaarachchi

Author 6: Sanjeeva Perera

Author 7: Chamath Randula

Keywords: Asset management; bulk asset transfer; Business Process Reengineering (BPR); Enterprise Resource Planning (ERP); workflow optimization

Paper 99: Stochastic Policies, Deterministic Minds: A Calibrated Evaluation Protocol and Diagnostics for Deep Reinforcement Learning

Abstract: Deep reinforcement learning (DRL) typically in-volves training agents with stochastic exploration policies while evaluating them deterministically. This discrepancy between stochastic training and deterministic evaluation introduces a potential objective mismatch, raising questions about the validity of current evaluation practices. Our study involved training 40 Proximal Policy Optimization agents across eight Atari environments and examined eleven evaluation policies ranging from deterministic to high-entropy strategies. We analyzed mean episode rewards and their coefficient of variation while assessing one-step temporal-difference errors related to low-confidence actions for value-function calibration. Our findings indicate that the optimal evaluation policy is highly dependent on the environment. deterministic evaluation performed best in three games, while low-to-moderate-entropy policies yielded higher returns in five, with a significant improvement of over 57% in Breakout. However, increased policy entropy generally degraded stability—evidenced by a rise in the coefficient of variation in Pong from 0.00 to 2.90. Additionally, low-confidence actions often revealed an over-optimistic value function, exemplified by negative TD errors, including -10.67 in KungFuMaster. We recommend treating evaluation-time entropy as a tunable hyperparameter, starting with deterministic or low-temperature softmax settings to optimize both return and stability on held-out seeds. These insights provide actionable strategies for practitioners aiming to enhance their DRL-based agents.

Author 1: Sooyoung Jang

Author 2: Seungho Yang

Author 3: Changbeom Choi

Keywords: Deep reinforcement learning; policy evaluation; stochastic policy; temporal difference error; Atari; PPO

Paper 100: Expert Systems in Tuberculosis Prevention Established in Certainty Factor

Abstract: Tuberculosis remains a highly relevant public health concern, especially in contexts with limited access to medical services, highlighting the need for tools that support early diagnosis. In this study, a web-based expert system was developed to assist in tuberculosis detection, using Buchanan’s methodology, which consists of five phases: identification, conceptualization, formalization, implementation, and validation. The system was designed with a knowledge rule-based approach and incorporated the Certainty Factor to quantify confidence in diagnostic conclusions. Validation was carried out through expert judgment using a 15-question survey. The results showed a high overall positive consensus, with question 13 standing out as it obtained the highest mean score (4.80) and the lowest dispersion (SD = 0.61), reflecting the most favorable perception and greatest agreement among the experts. Conversely, question 4 recorded the lowest mean score (4.00) and the highest dispersion (SD = 1.12), indicating aspects of the system that generated more divided opinions. Overall, these findings confirm that the system is effective, reliable, and usable, making it a relevant tool to support clinical decision-making in resource-limited settings. As an additional contribution, the integration of complementary technologies is suggested, such as (ML) algorithms, radiological image analysis, and mobile applications for symptom tracking, in order to optimize early detection and strengthen clinical care for tuberculosis.

Author 1: Inooc Rubio Paucar

Author 2: Cesar Yactayo-Arias

Author 3: Laberiano Andrade-Arenas

Keywords: Buchanan’s methodology; certainty factor; expert system; public health; tuberculosis; web application

Paper 101: Feature-Optimized Machine Learning for High-Accuracy Ammunition Detection in X-Ray Security Screening

Abstract: This paper introduces a machine learning system that is feature-optimized to enhance the detection of concealed ammunition in X-Ray security imaging. The system integrates advanced image analysis techniques with a cascade-AdaBoost classifier and Multi-scale Block Local Binary Pattern (MB-LBP) features, which are particularly effective for object recognition and classification in complex, high-dimensional data. The combination of these algorithms ensures robust performance in identifying ammunition types even under challenging conditions, such as variations in image quality or object orientation. The system is specifically designed for the accurate identification of various types of ammunition, including 9 mm bullets for handguns, AK-47 machine gun bullets, and 12-gauge shotgun cartridges. To support the development and testing of this system, a new dataset comprising 1,732 X-Ray images of passenger luggage was collected. This dataset is made publicly available to facilitate further research and improvement in this critical area of security technology. Experimental results demonstrate that the system achieves a high level of detection accuracy, with the ability to identify 12-gauge shotgun shells concealed in baggage with a 92% success rate. Beyond its technical achievements, this system significantly enhances the efficiency and reliability of security checks, improving the overall effectiveness of ammunition detection in real-world scenarios.

Author 1: Osama Dorgham

Author 2: Nijad Al-Najdawi

Author 3: Mohammad H. Ryalat

Author 4: Sara Tedmori

Author 5: Sanad Aburass

Keywords: Feature optimization; ammunition detection; X-Ray images; machine learning; security imaging

Paper 102: Evaluating Transparency in the Development of Artificial Intelligence Systems: A Systematic Literature Review

Abstract: Transparency is increasingly recognised as a cornerstone of trustworthy artificial intelligence (AI), yet its operationalisation remains fragmented and underdeveloped. Existing methods often rely on qualitative checklists or domain-specific case studies, limiting comparability, reproducibility, and regulatory alignment. This paper presents a Systematic Literature Review (SLR) of 28 peer-reviewed studies that explicitly propose or apply methods for evaluating transparency in AI systems (2019-July 2025). The review identifies recurring themes such as traceability, explainability, and communication, and classifies evaluation approaches by metric type and calculation type. Empirically, checklist-based instruments are the most frequent evaluation form (9/28, 32%), followed by scenario-based qualitative assessments (5/28, 18%). Most (9/28, 32%) research on AI applications occurs in healthcare; references to legal or ethical frameworks appear in 19/28 studies (67%), although traceable mappings to specific obligations are rare. The results of the quality assessment highlight strengths in methodological clarity, but reveal persistent gaps in benchmarking, stakeholder inclusion, and lifecycle integration. Based on these findings, this study informs the adaptation of the Z-Inspection® process within the context of AI development projects and motivates a Transparency Artefact Registry (TAR), a structured, metadata-based mechanism for capturing and reusing transparency artefacts across system lifecycles. By embedding transparency evaluation into AI development workflows, the proposed approach seeks to provide verifiable, repeatable, and regulation-aligned practices for assessing transparency in complex AI systems.

Author 1: Giulia Karanxha

Author 2: Paulinus Ofem

Keywords: Artificial intelligence; transparency evaluation; trustworthy AI; transparency metrics; EU AI Act; systematic literature review

Paper 103: Auditable Real-Time Cold-Chain Monitoring with IoT and Blockchain Anchoring

Abstract: Safe vaccine storage hinges on continuous, trust-worthy temperature supervision and evidence that records have not been altered. Yet many cold rooms still rely on fragmented logging tools that lack real-time alerts, end-to-end traceability, and audit-ready data. This paper presents a practical, low-cost architecture that integrates Internet of Things (IoT) sensing with blockchain anchoring to deliver real-time monitoring and visibility, reliable anomaly detection, and tamper-evident provenance for cold-chain storage. The design couples minute-level telemetry and dashboarding with alert debouncing/hysteresis to reduce false alarms, while anchoring hourly summaries and event alerts on-chain to create a verifiable trail without exposing raw data or incurring recurring fees during experimentation. A prototype in a vaccine cold-room scenario demonstrates that the approach is simple to deploy on commodity hardware, scales by adding room-s/sensors, and produces operator-friendly notifications along-side independently verifiable records. This combination of edge retention and cryptographic anchoring provides a pragmatic path for pharmacies, clinics, and warehouses to upgrade from basic loggers to transparent, audit-ready monitoring, bridging operational needs (alerts) and compliance needs (provenance) in one system.

Author 1: Mohamed DOUBIZ

Author 2: Mouad BANANE

Author 3: Abdelali ZAKRANI

Author 4: Allae ERRAISSI

Keywords: Internet of Things; blockchain; cold chain; vaccine storage; real-time monitoring; tamper-evident; data provenance; traceability

Paper 104: Systematic Literature Review of Reactive Jamming Attacks Mitigation Techniques in Internet of Things Networks

Abstract: Internet of Things (IoT) networks have become a prevalently exploited research area in academia and industry. IoT networks benefit from a variety of applications, including smart cities, smart homes, intelligent transportation, smart agriculture, monitoring, surveillance, etc. The security challenges associated with IoT networks have been broadly studied in the literature. This systematic literature review (SLR) is aimed at reviewing the existing research studies on IoT networks’ reactive jamming attacks, challenges, and mitigation. This SLR examined the research studies published between 2019 and 2024 within the popular electronic digital libraries. We selected 45 papers after a rigorous screening of published works to answer the proposed research questions. The outcomes of this SLR reported three major IoT network performance issues. The results showed that the existing mitigation methods are categorized as machine learning based, deception-based, statistical-based, radio frequency-based, game theory-based, and encryption-based. The results show that most methods can detect reactive jamming attacks with accuracy. However, those methods still require additional infrastructure, encryption systems, and lead to prolonged training delays due to large datasets, resulting in computational overhead and transmission delays. Furthermore, the methods are unable to provide a better defense response to reactive jamming attacks. This is because the methods cannot adequately deal with the increased power consumption of IoT devices, cannot minimize transmission delays, and cannot improve the packet delivery ratio. As a result, reactive jamming attacks continue to be prevalent in IoT networks.

Author 1: Enos Letsoalo

Author 2: Topside Mathonsi

Author 3: Tshimangadzo Tshilongamulenzhe

Author 4: Daniel du Duplesis

Keywords: IoT networks; reactive jamming attacks; mitigation methods; systematic literature review; electronic digital libraries

Paper 105: Virtual Assistant Based on Recurrent Neural Networks: Scope and Coverage of Health Insurance

Abstract: In Peru, the use of health services provided by state institutions has decreased, largely due to perceived deficiencies in care quality, such as delays in medical attention and administrative barriers that hinder timely access to information on insurance scope and coverage. This study develops a web-based application with an integrated chatbot that leverages Artificial Intelligence (AI) and a Recurrent Neural Network (RNN) to provide accurate and accessible information about institutional insurance, including the characteristics of different insurance types and specific coverage cases. The application integrates a chatbot powered by an AI model based on Natural Language Processing (NLP) from OpenAI, with intent recognition handled by a dedicated classifier. In comparison to existing chatbots in the healthcare field, the model proposes a hybrid approach based on RNN and GPT-3.5, optimized for health insurance queries in public sector contexts. The system was evaluated with 50 representative questions scored on four criteria: clarity, relevance, coherence, and accuracy. The chatbot achieved an overall mean response accuracy of 82% and a mean user satisfaction score of 4.59/5, indicating strong acceptance and usability. These results suggest that the combination of these technologies constitutes an effective alternative for addressing queries relevant to users of state health systems.

Author 1: Diego Alberto Paz-Medina

Author 2: David Eduardo Rojas-Cavassa

Author 3: Ernesto Adolfo Carrera-Salas

Keywords: Insurance; insurance coverage; web application; chatbot; Recurrent Neural Networks

Paper 106: Text Information Data Mining Method in Natural Language Processing Tasks

Abstract: Text mining methods often rely on a single data source or simple word frequency statistics, making it difficult to capture multi-source text semantic associations and local contextual dependencies, resulting in poor mining accuracy. Therefore, a method for text information data mining in natural language processing tasks is proposed. Using Python web crawlers to obtain multi-source text data, after preprocessing such as cleaning, segmentation, and removal of stop words, a Vector Space Model (VSM) is used for text representation, and a TF-IDF (Term Frequency Across Document Frequency) weight optimization mechanism is introduced to enhance feature semantic representation. On this basis, a semantic enhancement system is constructed based on the BERT classification model in the field of natural language processing. Through the self-attention mechanism of multi-layer Transformer encoders, semantics are aggregated to effectively capture local contextual dependencies, and context-sensitive word vectors are generated by the output layer. Finally, by fine-tuning the parameters of the BERT (Bidirectional Encoder Representation from Transformers) model and combining it with the Softmax function, precise mining of text information data categories was achieved. The experimental results show that in the embedding experiment of sports news headlines, this method can form a semantic aggregation structure with clear domain logic for word vectors; In the cross domain short text classification experiment, the overall accuracy of this method on the dataset reached 95.7%, which was 19.5% and 18.7% higher than the comparative methods, effectively solving the cross domain ambiguity problem in natural language processing.

Author 1: Shengguo Guo

Author 2: Dandan Xing

Keywords: Natural language processing; text information; data mining; VSM; TF-IDF; BERT

Paper 107: Document Similarity Detection for Project Development Using Fused Interactive Attention Mechanisms

Abstract: This study introduces a novel multi-feature fusion model aimed at improving text similarity calculation in scientific and technological projects. The primary objective is to enhance the accuracy and efficiency of assessing text similarities, particularly in evaluating originality and identifying duplications in project submissions. To overcome the limitations of traditional text similarity methods (e.g., Vector Space Models, Latent Dirichlet Allocation, and TF-IDF) in capturing complex semantic and structural features, a hybrid model is proposed. The model combines word embeddings (word2vec and cw2vec), a Bi-LSTM network, and a multi-perspective convolutional neural network (MP-CNN) for effective feature extraction. Additionally, a fusion attention mechanism and interactive attention are incorporated to improve the extraction of semantic, contextual, and structural information. Experimental evaluation on two benchmark datasets demonstrates that the proposed model achieves an average precision of 0.75, a recall of 0.71, and an F1-score of 0.73, outperforming traditional methods (LDA, TF-IDF, Word2vec+Cosine) and deep learning baselines (Siamese-LSTM, MP-CNN) by more than 10% on average. These results confirm that the proposed architecture effectively balances semantic relevance and structural integrity, yielding superior similarity detection performance. The integration of advanced deep learning components—Bi-LSTM, MP-CNN, and attention mechanisms—substantially improves both the accuracy and efficiency of similarity evaluation, providing a more reliable and objective approach for scientific project assessment.

Author 1: Chao Zhang

Author 2: Ying Zhang

Author 3: Gang Yang

Author 4: Fan Hu

Keywords: Text similarity; multi-feature fusion model; word2vec; cw2vec; MP-CNN; fusion attention mechanism; semantic extraction; project evaluation

Paper 108: A Comparison Between FAHP-TOPSIS and FAHP-FTOPSIS Methods for Selecting the Best Products for Home-Based Sellers: A Performance Analysis

Abstract: This study discusses common problems faced by home-based sellers in determining the right product ideas to sell. To overcome this problem, a method is needed that can help home-based sellers to choose the right product. Therefore, a decision support system using a multi-criteria decision-making technique with a hybrid approach was applied, which integrates the FAHP-TOPSIS and FAHP-FTOPSIS methods in the product selection process. The analysis results show that the FAHP-TOPSIS method is more effective in producing product rankings, with alternative A5948 ranking first with a score of 0.946. Meanwhile, the FAHP-FTOPSIS method also placed the same alternative in first place with a score of 0.679. The findings in the ranking analysis showed that the addition of fuzzy did not affect the rankings but did affect the score value of the alternatives. Sensitivity Analysis using Mean Absolute Deviation (MAD), Mean Square Error (MSE), and Spearman Correlation (SC) was conducted. FAHP-TOPSIS performed best at Weight 1 (MAD 89, MSE 18.486, SC 0.972) and excelled at Weight 3 (MAD 144, MSE 51.791, SC 0.997), though more volatile at other weights. Overall, at the base weight (Weight 1), TOPSIS shows the best ranking stability (low MAD/MSE, high SC), while with shifted weights (especially Weight 3), FTOPSIS better maintains ordering (SC ≈ 1) despite higher error at Weight 2. Practically, TOPSIS suits baseline scenarios; FTOPSIS is more robust under weight variations, with error variance control still necessary. These findings provide a practical guideline: use FAHP-FTOPSIS when preferences are uncertain, and use FAHP-TOPSIS when preferences are clear. The resulting rankings can be directly adopted by sellers to prioritize and select products with confidence.

Author 1: Selvia Lorena Br Ginting

Author 2: Zulaiha Ali Othman

Keywords: Home-based sellers; product selection; FAHP-TOPSIS; FAHP-FTOPSIS; sensitivity analysis