IJACSA Volume 17 Issue 6

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

View Full Issue

Paper 1: A Multi-Stage Framework for Bias Detection and Mitigation in AI-Driven Recruitment Systems

Abstract: The use of machine learning in recruitment has raised growing concerns about fairness, as automated hiring systems can generate unequal outcomes across demographic groups. These disparities are influenced not only by imbalanced data but also by the behavior of learning algorithms, making bias a multidimensional challenge that cannot be effectively addressed with single-stage solutions. This study introduces an integrated framework for bias detection and mitigation in AI-driven recruitment systems, combining interventions at the data, model, and decision levels within a unified evaluation pipeline. The framework is assessed using multiple classification models of varying complexity and evaluated with established fairness metrics. In addition, explainability techniques are employed using SHAP-based feature attribution to investigate hidden dependencies and assess the sensitivity of predictions to demographic attributes. Experimental results show that baseline models achieve strong predictive performance, with accuracy ranging from 0.807 to 0.816; however, fairness evaluation reveals substantial disparities, with Disparate Impact as low as 0.190 and Demographic Parity Difference exceeding 0.27 in some cases. After applying the proposed multi-stage mitigation approach, fairness metrics improve significantly: Disparate Impact meets or exceeds the legal threshold of 0.80 across all models, with reductions in Demographic Parity Difference of 70–85% and Equal Opportunity Difference of 47–75%, and demographic disparities are reduced to 0.029–0.053. These improvements are achieved with minimal performance trade-off, as overall accuracy decreases by at most 4.2 % points while ROC AUC remains unchanged. The findings demonstrate that bias in recruitment systems arises from the interplay between data and model dynamics and highlight the importance of coordinated mitigation strategies throughout the machine learning lifecycle. This work provides a practical, scalable approach to developing fair and transparent AI systems for hiring applications.

Author 1: Gideon Assafuah

Author 2: Claude Turner

Author 3: Carlene Turner

Author 4: Kingsley Nwosu

Keywords: Bias detection; bias mitigation; AI in recruitment; algorithmic fairness; explainable AI; fairness metrics; machine learning; responsible AI

Paper 2: Zero-Disclosure Material Passports for Verifiable Provenance in Multi-Tier Supply Networks

Abstract: Global supply chains produce vast quantities of transactional data, yet most existing traceability systems force companies to choose between disclosing sensitive commercial relationships to a shared infrastructure and relying on mech-anisms that do not provide strong privacy guarantees. This study introduces a zero-disclosure material passport framework for verifiable provenance in multi-tier supply networks. The framework represents product history as a Resource-Event-Agent provenance graph and authenticates each event through unlinkable aggregate signatures. Individual participant signatures collapse into a single, fixed-size aggregate proof, so the crypto-graphic verification object remains 192 bytes in size regardless of the supply chain depth. Selective disclosure is supported through commitment-based predicate proofs that allow regulators and auditors to verify attributes such as origin, certification status, or recycled-content thresholds without learning unrelated commercial information. A Rust prototype was evaluated on an x86-64 workstation, a Raspberry Pi 4, and a Raspberry Pi Zero 2 W. Signing latency remained below 700 ms on the most constrained device, demonstrating feasibility for low-frequency supply chain handover events and offline inspection contexts. The study further provides algorithmic descriptions of credential issuance, passport update, aggregate verification, and selective disclosure; an expanded comparison against blockchain, EDI, BBS+/anonymous-credential, and zk-SNARK-based approaches; and a discussion of the centralized Credential Authority as an explicit trust assumption. Security analysis establishes unforge-ability and unlinkability under standard hardness assumptions when the Credential Authority is honest, and participant creden-tials are issued through the prescribed blinded protocol.

Author 1: Shivani Dharmavaram

Keywords: Supply chain traceability; material passports; zero-knowledge proofs; aggregate signatures; privacy-preserving authentication; circular economy; provenance verification; selective disclosure

Paper 3: Exploring the Structure Resulting from Unstructured Neural Network Pruning

Abstract: Iterative Magnitude Pruning (IMP) is a widely used technique for compressing neural networks by progressively removing low-magnitude weights while maintaining predictive accuracy. Despite its widespread application and simplicity, the underlying reasons for its effectiveness remain underexplored. In this work, the pruning dynamics and emergent structural traits of IMP are empirically examined across three benchmark convolutional architectures (ResNet20, Vgg16, and RegNetX) and evaluated on the CIFAR10 and Tiny ImageNet datasets. A comprehensive analysis reveals that IMP preferentially removes weights from deeper layers, preserving early feature extractors until a critical network sparsity threshold of 96% is reached. Up to this 96% threshold, test accuracy remains remarkably stable across all evaluated networks before experiencing significant degradation. Quantitative evidence is provided showing that later stages are pruned more heavily, taking advantage of the fact that deeper layers contain more low-magnitude weights. Furthermore, neuron-level investigations reveal that IMP does not produce any completely pruned (100% sparse) neurons, even at extreme sparsity levels. Sensitivity analysis via neuron zeroing demonstrates that individual neurons maintain a stable range of functional importance rather than narrowing as pruning progresses. Finally, activation similarity metrics indicate that feature representations are preserved throughout the pruning process, keeping pruned and unpruned networks in close rep-resentational alignment. These findings highlight the surprising degree of structure generated by unstructured pruning, offering new insights into network compression and potential pathways for hardware-efficient sparse matrix adaptations.

Author 1: Jamil Gafur

Author 2: Max Milkert

Author 3: Kevin Patrick Griffin

Author 4: Nicholas T. Wimer

Author 5: Charles Tripp

Author 6: Steve Goddard

Keywords: Machine learning; convolutional neural networks; multi-layer perceptrons; artificial intelligence

Paper 4: Privacy Leakage and Memorization in Fine-Tuned Clinical Language Models: A Controlled Study of Defenses and Backbone Choice on Clinical Narrative Transcriptions

Abstract: The increasing adoption of large language mod-els (LLMs) and domain-adapted transformers in healthcare has created a new privacy challenge: fine-tuned models may memorize rare clinical strings and later reveal them through generation or scoring behavior. A controlled study of privacy leakage and memorization in clinical language models trained on narrative transcriptions is presented. A canary-based audit pipeline was instantiated on a 4,000-note subset of the Medical Transcriptions (MTSamples) corpus, with 40 synthetic secrets injected only into the training partition and evaluated using three complementary attack families: prompt extraction, exposure-style ranking, and reference-based membership inference. Two experiments are reported. Experiment I compares baseline fine-tuning, early stopping, and a conservative regularized training profile combining lower learning rate, higher weight decay, and partial layer freezing. Experiment II fixes the training protocol and compares DistilGPT2, GPT-2, and BioGPT. A clear privacy-utility tension was observed. In Experiment I, early stopping produced the best held-out language-model utility, whereas the combined regularized profile eliminated observed prompt leak-age and reduced membership-inference strength, at the cost of worse perplexity. In Experiment II, stronger and more domain-specialized backbones achieved better clinical language modeling but also exhibited higher leakage and stronger membership-inference signals, with BioGPT yielding the strongest utility and the highest privacy risk under the evaluated attacks. These results indicate that privacy auditing should accompany utility evaluation in clinical LLM adaptation, and that backbone choice can materially affect memorization risk in this controlled setting.

Author 1: Yassine Chahid

Author 2: Anas Chahid

Author 3: Ismail Chahid

Author 4: Aissa Kerkour Elmiad

Keywords: Clinical language models; privacy leakage; memo-rization; membership inference; canary exposure; BioGPT; GPT-2; healthcare NLP; model auditing

Paper 5: Deployer-Side XAI Instrumentation for Regulated AI: A Clinical Case Study in ICL Sizing

Abstract: Regulated AI creates a monitoring problem for deployers who must organise human oversight, log-retention and post-market surveillance while often having access only to the prediction interface. This study specifies a deployer-side XAI instrumentation protocol for the output→action boundary, where a model output becomes a reason for action. The protocol reorganises KernelSHAP, nearest-neighbour envelope checks, bounded perturbation, and rank-order stability into a per-decision evidence record computed from predict() calls. We instantiate the protocol in a clinical case of phakic Implantable Collamer Lens sizing, using a 55-eye held-out cohort and an Extra Trees regressor for post-operative vault prediction. The record contains five signals: score_margin, constraint_enforcement, envelope_validity, decision_robustness, and record_integrity, plus two cohort-level oversight aggregates. The case study shows how the same record can support decision-time human oversight, later audit and post-market surveillance under the EU AI Act and the Medical Devices Regulation.

Author 1: Dorleta Urrutia-Onate

Author 2: Enrique Onieva

Author 3: Asier Perallos

Keywords: Explainable AI; XAI instrumentation; human over-sight; Medical Device Software; EU AI Act; ICL sizing

Paper 6: DcDM: A Pre-training Data Evaluation Framework for Proactive Drift Prevention in Machine Learning

Abstract: The performance of machine learning (ML) systems often deteriorates over time owing to data drift, which is typically caused by changes in data quality or distribution. Such degradation in deployment environments can result in inaccurate predictions and reduced system reliability. Conventional drift detection approaches have largely focused on retraining ML models after performance degradation has occurred. However, because the root causes of drift often originate from the data itself, a data-centric approach is needed to address the problem at its source. This study proposes a systematic, data-centric drift management (DcDM) framework that integrates domain-specific rule validation, data quality assessment, and statistical drift analysis before model training. By first verifying semantic constraints and then assessing data quality and distributional stability, DcDM enables early identification of potential drift hazards and prevents low-quality or semantically invalid data from entering the training pipeline. We evaluate the proposed framework on three datasets across different modalities: Electricity Load Diagrams, CIFAR-10, and VisDA-2017. The experimental results show consistent improvements of approximately 6% to 12.5% in both accuracy and F1-score. Additional ablation studies, baseline comparisons, and statistical significance tests further demonstrate the robustness and effectiveness of the proposed approach.

Author 1: Okjoo Choi

Author 2: Wonsun Shin

Keywords: Data drift; data evaluation; data quality; domain rule; drift prevention and mitigation; drift management

Paper 7: Characterizing Operational Drift in Cement Manufacturing Process Data via Self-Supervised Representation Learning

Abstract: Cement finish milling generates large volumes of process variable data at hourly resolution, while quality measurements (Blaine fineness, 44 µm residue) are recorded only every two to four hours, producing a sparse-label regime with substantial unlabeled data accumulated over multi-year operation. In this study, we analyze 188,858 hourly records collected from three industrial cement mill units spanning 2017 to 2025 (approximately nine years) and characterize the operational drift structure embedded in the data. Using Pruned Exact Linear Time (PELT) change-point detection on three critical process variables, we identify between 12 and 21 detected drift events per mill, with Cohen's d effect sizes exceeding 1.0 for the majority of detected change points and reaching 4.3 in extreme cases. We then apply SCARF, a contrastive self-supervised learning method for tabular data, to learn 128-dimensional representations on the combined labeled (51,225 records) and unlabeled (111,506 records) data. Multi-seed training yields stable validation InfoNCE loss of 6.93 ± 0.02. Three clustering algorithms applied to the learned embedding space (K-means, Gaussian Mixture Model, HDBSCAN) consistently select 15 to 21 operational regimes, with silhouette scores between 0.36 and 0.41. The Adjusted Rand Index between embedding-space K-means and process variable space K-means is 0.35, indicating that the learned representation preserves coarse regime structure while resolving finer sub-regime variability. Cluster analysis further reveals strong mill specificity, with 14 of 15 embedding clusters dominated by a single mill, and temporal cluster evolution that aligns with PELT-detected change-point boundaries. These findings establish that long-term cement process data contains a richer operational regime structure than implied by raw process variable clustering, and that self-supervised pretraining can recover this structure, and that the resulting representation yields statistically significant gains in downstream quality prediction.

Author 1: Changgyun Kim

Keywords: Cement manufacturing; self-supervised learning; change-point detection; operational drift; tabular data; representation learning; SCARF; PELT

Paper 8: Digital-to-Physical Transfer of Adversarial Patches for Aerial Vehicle Detection

Abstract: Deep neural network (DNN)-based object detec-tors are widely used for analyzing aerial and satellite imagery in applications such as environmental monitoring and urban analytics. Despite their strong performance, these models are known to be vulnerable to adversarial examples, and physical adversarial attacks using printable patterns pose realistic security threats. This study evaluates physical adversarial patch attacks against an aerial vehicle detector by bridging digital optimization and real-world deployment. Adversarial patches are optimized in the digital domain using a loss function that minimizes the maximum objectness score while incorporating non-printability score (NPS) and total variation (TV) constraints to ensure both printability and spatial smoothness. The optimized patches are printed and deployed in three configurations: ON, OFF, and OFF-Side. Experiments using a YOLOv3 detector show that while the OFF patch achieves the highest effectiveness in the digital domain (85.51% Average Objectness Reduction Rate (AORR)), the ON patch demonstrates superior robustness in physical environments (0.197–0.343 Objectness Score Ratio (OSR)) due to its consistent visibility. Furthermore, the results indicate that weather-based augmentation does not necessarily improve patch optimization in this domain. These findings provide critical insights into the practical vulnerabilities of aerial object detection systems.

Author 1: Jung Heum Woo

Author 2: Eun-Kyu Lee

Keywords: Aerial object detection; physical adversarial attack; adversarial patch; patch optimization; security; attack; feature

Paper 9: A Hybrid Ethereum-Based Architecture for Secure Electronic Health Records: Consent, Integrity Anchoring and Auditable Access

Abstract: Securing electronic health records (EHR) requires strong guarantees for confidentiality, integrity, access control, and auditability. Traditional centralized architectures rely on database-level protection and internal logging, which remain vulnerable to insider misuse and undetected data modification. This study proposes a practical hybrid architecture in which medical content is stored encrypted off-chain, while blockchain is used selectively as a governance and evidence layer. An Ethereum-based prototype was designed and implemented to support integrity anchoring of medical documents, patient-controlled consent management, and immutable audit trails for critical actions. In the implemented solution, the actual medical content is not stored on-chain. Instead, the blockchain stores only document-related metadata, cryptographic hashes, document references, and access-control information, while the sensitive medical data remains encrypted and stored off-chain. This design supports GDPR-oriented data minimization, since the immutable blockchain layer does not contain raw medical records or directly identifiable medical content. The prototype separates confidentiality from blockchain immutability. Medical document confidentiality is handled at the application and off-chain storage level, while the blockchain is used for integrity verification, consent management, and auditability. Encryption keys are not stored on-chain, which prevents the blockchain layer from becoming a repository of sensitive or directly exploitable medical information. Security mechanisms are integrated directly into application flows, including hash-based tamper detection and on-chain verification of access rights. The prototype is evaluated through realistic operational scenarios, analyzing security properties, performance, and transaction cost implications. Results show that, relative to a DB-only baseline, the hybrid approach provides structurally stronger support for integrity verification, traceability, and accountability without exposing sensitive medical data on-chain. The study also highlights practical limitations related to latency and costs in public blockchain environments, supporting a selective on-chain design focused on high-value operations.

Author 1: Rodica Doina Zmaranda

Author 2: Attila-Imre Kovacs

Author 3: Daniela Elena Popescu

Author 4: Alexandrina Mirela Pater

Keywords: Electronic Health Records (EHR); blockchain; ethereum; smart contracts; access control; consent management; auditable access

Paper 10: Using Artificial Intelligence Techniques for Error Detection and Repair in Software Development

Abstract: Software bugs remain one of the most costly chal-lenge in software engineering, consuming significant development time and resources. Recent advances in Artificial Intelligence (AI), particularly deep learning and large language models (LLMs), have shown remarkable potential in automating the detection and repair of software errors. This study presents a comprehensive survey and comparative analysis of AI-based techniques for error detection and automated program repair (APR). We catego-rize existing approaches into traditional search-based methods, learning-based neural machine translation models, and emerg-ing LLM-based repair systems. We evaluate these techniques across standard benchmarks, including Defects4J and SWE-bench, comparing their effectiveness in terms of bugs fixed, patch correctness, and scalability. Our analysis reveals that while LLM-based approaches significantly outperform traditional methods in repair capability, challenges remain in patch correctness validation, computational cost, and generalization to real-world codebases. Our results show that LLM-based tools, such as ChatRepair, can correctly fix 114 out of 395 benchmark bugs at just $0.42 per fix, fixing 2.6× as many bugs as the best traditional method (a 165% increase). We discuss open challenges and propose future research directions toward more reliable AI-assisted software development.

Author 1: Abdulaziz Aladwani

Author 2: Sultan Alsamaani

Author 3: Turki Alrumaykhani

Author 4: Mohamed Tahar Ben Othman

Keywords: Automated program repair; bug detection; deep learning; large language models; software engineering; neural machine translation; defect prediction; code analysis

Paper 11: Analyzing User Experience in Mobile Banking Applications Through Text Mining

Abstract: This study attempts to offer a data-driven comprehension of the factors that influence satisfaction and dissatisfaction by examining user reviews, based on a banking mobile application in Albania, specifically the Raiffeisen Bank application. All gathered reviews from Google Play and the App Store have been classified using the multilingual BERT model. It is important to mention that mBERT is particularly suitable given that the reviews are written in both Albanian and English. BERTopic has been used to determine the main topics of classification. After classifying the reviews and creating new dimensions with BERTopic, an ordinal logistic regression model was applied to the generated dataset to assess the predictive power of these dimensions and sentiment polarity on user-assigned satisfaction ratings. This study helps developers and managers to better understand the main factors that influence customer reactions and evaluate the factors that influence the rate of the application and the different bugs they report.

Author 1: Bora Lamaj (Myrto)

Author 2: Markela Muça

Author 3: Klodiana Bani

Keywords: Generated reviews; sentiment analysis; topic modeling; BERTopic; mBERT

Paper 12: Design and Evaluation of a 13K Ultra-High-Resolution Web-Based Virtual Reality Platform for Immersive Spatial Visualization

Abstract: Delivering ultra-high-resolution immersive environments via standard web browsers presents significant rendering and architectural challenges. While Virtual Reality (VR) is widely adopted, its application as a high-fidelity spatial visualization tool often lacks robust empirical evaluation regarding user experience and system efficacy. This study proposes the design, development, and evaluation of a 13K ultra-high-resolution web-based virtual tour platform, utilizing a nature-based destination as a practical case study to assess system effectiveness. The system architecture was engineered to process and render 13K panoramic data with interactive navigational hotspots, ensuring cross-platform web compatibility. To evaluate the Human-Computer Interaction performance, a true experimental design was employed with eighty participants who were randomly assigned to two groups. The experimental group interacted with the proposed 13K web-based platform, while the control group utilized a conventional static two-dimensional web interface. System efficacy was measured through user-centric metrics, including Sense of Presence, spatial image perception, and subsequent behavioral intention. The evaluation results demonstrated a statistically significant superiority of the proposed 13K platform. Users reported a substantially higher Sense of Presence score of 4.35 out of 5, compared to the conventional web interface score of 2.45. Furthermore, the immersive system architecture significantly catalyzed positive spatial perception and behavioral engagement, yielding a remarkably large effect size. The findings validate that the proposed high-fidelity web-based VR architecture effectively overcomes physical barriers, serving as a highly effective information system solution for remote spatial visualization and practical destination management.

Author 1: Worapon Manosroi

Author 2: Apisak Phromfaiy

Author 3: Pitak Khlaichom

Keywords: Virtual Reality; 360-degree panorama; web-based platform; immersive spatial visualization; ultra-high-resolution panorama; HCI (Human-Computer Interaction)

Paper 13: Artificial Intelligence in Software System Quality Assurance: A Systematic Literature Review of Techniques, Tools, and Challenges (2016–2026)

Abstract: A shift from deterministic testing to artificial intelligence (AI) driven quality ecosystems is necessitated by the rapid evolution of software architectures, and research from 2016 to the present year is combined by this systematic literature review (SLR), so 195 main studies are analyzed, and the path of artificial intelligence in Software Quality Assurance (SQA) is mapped. An enormous course in scholarly output from 2023 onwards is revealed by the findings, and this growth is driven by the industrial adoption of Large Language Models (LLMs) alongside autonomous agentic systems. In addition, three main areas are addressed by this study, and these areas are identified as the taxonomic shift toward multi-agent architectures, the functional effect of AI on labor-intensive activities such as regression testing, and self-healing automation, and the emerging social and technical challenges of ethical governance alongside explainability. Despite incomparable efficiency gains being offered by AI-driven techniques, the industrial success of these tools is strictly limited until the Maintenance Crisis of generated code is resolved and transparency is ensured through Explainable AI (XAI). Finally, the study concludes with a strategic roadmap for Ethical SQA, providing a foundation for future research in autonomous, self-evolving software systems.

Author 1: Abdullah A H Alzahrani

Keywords: Software Quality Assurance (SQA); artificial intelligence (AI); systematic literature review (SLR); Large Language Models (LLMs); Explainable AI (XAI)

Paper 14: A Comparative Evaluation of Large Language Models for Named Entity Recognition in Cyber Threat Intelligence

Abstract: Cyber Threat Intelligence reports combine analytical prose with dense technical indicators, making structured entity extraction a challenging but operationally valuable task. This study presents a comparative evaluation of three large language models – Claude Sonnet 4.6, GPT-5.4, and LLaMA 4 Scout – on a manu-ally annotated corpus of 21 real-world CTI reports across 15 entity types and 1284 ground truth instances. This study evalu-ates zero-shot and few-shot prompting conditions and studies the effect of iterative prompt refinement, focusing on explicit format constraints for cryptographic hash entities. Results show that Claude Sonnet 4.6 and GPT-5.4 achieve comparable perfor-mance under zero-shot conditions, with LLaMA 4 Scout trailing by a substantial margin. Few-shot prompting consistently reduc-es hallucination rates, but yields mixed F1 results, with exemplar cardinality emerging as a critical and underappreciated design factor. Entity extraction difficulty varies substantially across types, with technical indicator categories showing near-perfect performance and semantic categories such as tool and target sector posing the greatest challenges across all evaluated mod-els.

Author 1: Aykhan Huseynli

Keywords: Cyber threat intelligence; named entity recognition; large lan-guage models; prompt engineering

Paper 15: A QR-Code-Based Mobile Learning System for Science Instruction in Resource-Limited Schools

Abstract: This study developed and evaluated a QR-Code-Based Mobile Learning System for Science Instruction in Resource-Limited Schools. The system was designed to provide junior high school students with low-cost, mobile-accessible science learning materials through printed QR Code Cards linked to a mobile-responsive digital repository. The developed module set consisted of six curriculum-aligned Life Science topics: Cell: The Basic Unit of Life, Cell Theory, Cell Structure and Functions, Prokaryotic vs. Eukaryotic Cells, Cell Types and Cell Modifications, and Cell Cycle. A developmental, quasi-experimental, and descriptive-evaluative research design was employed. The participants were 80 Grade 9 and Grade 10 students divided into an experimental group using the QR-code-based mobile learning system and a control group using traditional printed flashcards. Results showed that students had moderate overall familiarity with QR codes (M = 3.52), highly positive attitudes toward the system (M = 4.45), and high usability ratings (M = 4.38). The QR Code Cards also received very high ratings in terms of portability, durability, readability, and sustainability (M = 4.50). Critically, the experimental group obtained a post-test mean score of 89.15 compared with the control group's mean score of 78.42, yielding a statistically significant difference (t = 6.24, p = .002, Cohen's d = 2.49, 95% CI [7.17, 14.29]), indicating that the QR-code-based mobile learning system significantly improved science achievement by a mean margin of 10.73 points. The findings suggest that QR-code-supported mobile learning can serve as a practical, scalable, and cost-effective instructional innovation for improving science instruction in resource-limited school settings.

Author 1: Gina B. Selga

Keywords: Academic performance; educational technology; mobile learning; QR code; resource-limited schools; science instruction; usability

Paper 16: An Improved Pre-processing Method for High-Quality MRI Images in Brain Tumor Detection

Abstract: Magnetic Resonance Imaging (MRI) is among the effective methodologies to identify tumors in the brain, but this method may not be very reliable because of the challenges in acquiring the images, which may include image noise, contrast differences, and spatial variation of intensity. To solve these problems, this study suggests an innovative pre-processing framework that will be used to improve the quality of MRI images prior to segmentation and tumor analysis. This study critically compares some noise removal methods, such as Gaussian, median, Wiener, and guided filtering, as well as Discrete Wavelet Transform (DWT)-based de-noising with soft thresholding. The new hybrid model, based on the combination of the advantages of various methods, is a Wavelet-NLM-Median (WNM). WNM integrates multiresolution wavelet shrinkage, non-local redundancy modelling, and median-based edge preservation to achieve improved noise reduction while maintaining structural details. Extensive testing is performed across noise levels ranging from 5% to 50%, and performance is assessed using standard evaluation metrics such as PSNR, MSE, SSIM, and SNR. The proposed WNM hybrid model demonstrates the highest reconstruction quality at 5% noise, with a PSNR of 45.98 dB, MSE of 3.35, SSIM of 0.985, and SNR of 44.89 dB. These statistics were substantially better than those of the Wiener and Guided filters, as well as soft-thresholding based on DWT. Visual assessments also showed that the WNM hybrid approach does a better job of keeping tumor boundaries, fine textures, and structural patterns than any other baseline filtering method. This shows that it is better at restoring high-quality MRI images for later study. The improved MRI inputs used to pre-process training images for a deep learning segmentation model improve the accuracy of the segmentation and the sharpness of the boundaries in a way that could be quantified. The suggested WNM pipeline is quick, works with many types of modalities, and is simple to connect to clinical CAD systems. It is a giant leap in pre-processing the MRI to detect malignancies in the brain.

Author 1: Nirmala

Author 2: Kavitha B C

Keywords: Brain MRI de-noising; Wavelet–NLM–Median (WNM) model; structural similarity preservation; image quality enhancement

Paper 17: Vibration-Aided Picture-Based Authentication for Shoulder-Surfing Resistant Mobile Login

Abstract: Text and graphical passwords on smartphones are easy to shoulder-surf in public, and many of the alternatives that have been proposed do not work well on small touchscreens. We describe a vibration-aided picture-password scheme that pairs an image-based credential with hidden haptic prompts. The user selects an image during registration, picks at least three secret locations on it, defines a set of deceiving locations, and maps vibration patterns to specific decoy tap styles. The image, locations, vibration mapping, spatial acceptance regions, and timing thresholds are stored on a server. At login time, the server builds a per-session challenge that interleaves the secret and deceiving locations, assigns vibration patterns only to the decoys, and pushes the challenge to the client. The device displays the image, vibrates before each decoy tap, and records tap coordinates and durations. The server matches the input against the challenge, accepts or rejects accordingly, and regenerates a fresh challenge after any mismatch, so that replay and pattern-learning attempts gain nothing. Because the vibration cue is tactile and not visible from outside, an onlooker cannot tell a genuine tap from a decoy response. The result is better resistance to shoulder-surfing without any extra hardware and without asking the user to learn anything especially complex.

Author 1: Ibrahim Albadi

Author 2: Mahdi Almalki

Author 3: Abdullah Albokhari

Author 4: Salman Alsumairi

Author 5: Sami Atiyah

Author 6: Faisal Alsubaei

Author 7: Abdullah Abuhussein

Keywords: Picture-based authentication; vibration cues; shoulder-surfing; mobile security; haptics

Paper 18: AI-Based Career Transition Recommendation System Using Controlled Educational Data Augmentation

Abstract: Career transition into digital professions is a strategic lever for addressing youth unemployment in Sub-Saharan Africa. However, existing training programs lack objective career guidance tools. This study presents two complementary contributions, evaluated using data collected from 131 professionals who underwent a career transition process into a digital profession. Learn-Orient is a career path recommendation system combining a profile recognition filter based on Gower distance normalized by the IQR and a binary logistic regression model, producing probabilistic estimates with a graded confidence level (High, Moderate, Low). EDU-CDA is a data augmentation method suited for small educational datasets with mixed variables and imbalanced classes, combining conditional SMOTE for continuous variables and conditional Bernoulli sampling for binary variables. Validated by a three-way protocol (distributional overlap 82.9–90.8%, TVD < 0.061, TSTR Δ = 0.025), EDU-CDA transforms the logistic regression model, which is typically the most vulnerable to small sample sizes (AUC = 0.810 in real cross-validation)—into the most stable and high-performing one (AUC = 0.995 in cross-validation, F1 = 0.900 on a real test set of 20 observations). Among the 16 pre-training predictors selected, institutional funding (OR = 32.40) proves to be the most discriminating factor for the Developer profile, while the creativity test score (OR = 0.107) strongly characterizes the Designer profile. The system incorporates a filter for detecting atypical profiles, ensuring responsible use in a decision-making context with significant human stakes.

Author 1: Gerlix ADANKON

Author 2: Pelagie HOUNGUE

Author 3: Melckior DEGBOE

Author 4: Corelle GOGAN

Keywords: Machine learning; data augmentation; predictive model; artificial intelligence in education; professional retraining; career recommendation

Paper 19: Verifiable Learned PR-Tree Indexing for Privacy-Preserving Range Queries Over Encrypted Geospatial Data

Abstract: Protecting the privacy of geospatial data, along with efficient encrypted query processing, remains a major challenge in cloud-based GIS applications and location-based applications (LBS). In this study, a privacy-preserving framework for secure range query processing over encrypted vector geospatial data using a learned PR-tree index is integrated with an XGBoost-based bucket prediction model. In the first phase, the framework employs a lightweight dual-encryption scheme based on Lorentz and Galilean transformations. This encryption preserves coordinate relationships and enables reversible coordinate recovery. To improve query execution efficiency in the encrypted domain, the learned PR-tree predicts the most probable PR-tree buckets. This minimizes unnecessary search path traversal. Further, integrity verification during storage and query processing is ensured using the Merkle Hash Root and EdDSA digital signatures. Experimental evaluation was conducted using various real-world point, polyline, and polygon datasets. The proposed framework achieved prediction accuracy up to 97.2% with a very low mean bucket prediction error of 0.028. The obtained results demonstrate the efficiency and practicality of the proposed framework over encrypted cloud data.

Author 1: Anagha Aher

Author 2: Sangita Chaudhari

Keywords: Location privacy preservation; learned PR-tree indexing; range query processing; Merkle Hash Root verification; XGBoost for spatial prediction

Paper 20: A Multi-Modal Deep Learning Framework for Student Soft Skills Development in Adaptive Learning Environments

Abstract: The rapid evolution of artificial intelligence and adaptive educational technologies has created increasing demand for intelligent systems capable of automatically assessing and enhancing student soft skills within digital learning environments. This study proposes MST-SoftNet, a multimodal transformer-based deep learning framework designed for adaptive soft skill assessment and personalized educational recommendation. The proposed architecture integrates heterogeneous educational modalities, including textual interactions, speech signals, facial expressions, behavioral engagement patterns, performance indicators, and feedback information, within a unified hierarchical transformer fusion framework. Modality-specific encoders, cross-modal attention mechanisms, explainable attention visualization modules, and adaptive recommendation components were incorporated to improve both predictive performance and interpretability. Experimental evaluation was conducted using multiple baseline deep learning architectures, including CNN, LSTM, Transformer, and multimodal CNN-LSTM models. The proposed MST-SoftNet framework achieved superior performance across all evaluation metrics, attaining 93.67% accuracy, 92.11% F1-score, and 96.05% AUC, while simultaneously demonstrating reduced inference latency and improved computational efficiency. Attention visualization analysis further confirmed the capability of the framework to identify semantically meaningful multimodal behavioral patterns associated with communication, collaboration, leadership, creativity, emotional intelligence, and self-regulation competencies. Longitudinal adaptive learning experiments additionally demonstrated substantial improvement in student soft skill progression over time. The obtained results indicate that MST-SoftNet establishes a robust, interpretable, and scalable foundation for next-generation intelligent educational systems focused on personalized soft skill development and adaptive learning optimization.

Author 1: Marzhan Bekbolat

Author 2: Kamalbek Berkimbayev

Author 3: Rustam Abdrakhmanov

Author 4: Serik Kenesbayev

Author 5: Nuraim Ibragimova

Author 6: Zhaksylyk Dzhanabayev

Keywords: Deep learning; transformer architecture; soft skill assessment; adaptive learning environments; personalized learning; student behavior analysis; multimodal fusion

Paper 21: A Hybrid Multi-Objective AI Framework for Curriculum-Aware Examination Generation

Abstract: Automated examination generation has become increasingly important in modern education, where assessments must satisfy multiple pedagogical constraints, including cognitive balance, curriculum alignment, and question diversity. Existing approaches often address these requirements independently, limiting exam coherence and overall quality. To overcome this limitation, this study proposes a curriculum-aware hybrid framework that integrates Curriculum Knowledge Graphs (CKG), NSGA-II, and Proximal Policy Optimization (PPO). The problem is formulated as a constrained multi-objective optimization task that simultaneously maximizes Bloom’s taxonomy alignment, difficulty balance, and CLO coverage while minimizing semantic redundancy. The CKG captures relationships among questions, concepts, and CLOs to ensure structured curriculum alignment; NSGA-II generates Pareto-optimal exam candidates, and PPO further refines them through adaptive policy learning. The framework was evaluated on a dataset of 8,000 annotated questions using cross-validation, ablation studies, and statistical significance testing. Results demonstrate strong performance, achieving a Bloom Balance Score of 0.84 and CLO coverage of 87.5%, while reducing semantic redundancy from 10.7% to 3.2% (Δ = 7.5 percentage points; 70 % relative reduction, p < 0.001).

Author 1: Mohamed Fathy Yehia

Author 2: Yehia M. Helmi

Author 3: Mahmoud Mohamed Bahloul

Keywords: NSGA-II; Reinforcement learning; automated exam generation; Bloom’s taxonomy; CLO alignment

Paper 22: An Intelligent Scheduling Optimization Algorithm for Multimodal Cache Resources with Status Awareness of Metropolitan Area Network CDN Nodes

Abstract: To address the resource scheduling challenges faced by metropolitan area network content delivery networks (CDN) when carrying multimodal traffic streams such as high-definition video, virtual reality (VR), and augmented reality (AR), this study proposes an intelligent optimization algorithm for multimodal cache resource scheduling that is CDN Node State Awareness. First, to address the dynamic nature of network topology and the heterogeneity of service streams, we construct a directed graph-based metropolitan area network CDN model. This model enables real-time perception of multi-dimensional states of nodes, including CPU utilization, memory usage, remaining bandwidth, and cache occupancy. We also introduce a mechanism for quantifying the transmission demand weight and cache value of multimodal content, providing a foundational support for differentiated scheduling. Second, at the optimization enhancement layer, we design a transmission path selection strategy, a cache replacement mechanism that integrates content value and access popularity, and an adaptive scheduling structure based on node load balancing. Furthermore, a Deep Q-Network is introduced at the cloud computing decision layer. Node states and user request features are modeled as a state space, while cache placement and request allocation strategies are modeled as an action space. A multi-objective reward function integrating hit rate, response latency, and packet loss rate is designed to achieve dynamic and intelligent scheduling of multimodal cache resources. Integrating path selection, cache updates, and fault recovery mechanisms to construct an overall optimization model enhances the system's adaptive scheduling capability in complex business services. The experiment shows that the algorithm has significant advantages in multi node collaborative scheduling: within 1-8 seconds, the transmission rate of node B reaches 30Mbps and the resource utilization rate of node A is improved; The resource download time remains stable at 4.4-4.9 seconds during 24-hour operation; In the large-scale scenario of 500 user requests, cross node cache load adaptive balancing, system overhead linearly increases, and data transmission security rate reaches 99.45%, creating an efficient, reliable, and scalable intelligent scheduling system for multi-mode content distribution in metropolitan area networks.

Author 1: Ruirong Jiang

Author 2: Zhibiao Xiong

Author 3: Junliang Wu

Author 4: Jinyong Xu

Keywords: Metropolitan area network; CDN node state awareness; multimodal; cached resources; intelligent scheduling optimization; deep Q-Network; reward function

Paper 23: MSE-Guided Hybrid U-Net Framework for Automatic Kidney Segmentation and Spatial Localization

Abstract: Evaluating segmentation results in ultrasound imaging is still difficult due to noise, low contrast, and ambiguity at the boundaries, which makes it very challenging to measure accurately. Mean Squared Error (MSE) is a widely used but highly spatially sensitive evaluation metric for comparing predicted masks and ground truth. This work introduces a Mean Squared Error (MSE) based evaluation framework augmented using Block-Based Region Matching (BBRM) to achieve higher robustness against positional errors. The MSE is calculated under spatial shifts, and the best alignment with the lowest error is identified. To verify the effectiveness of the method, this work uses multiple deep learning segmentation models as baseline methods, along with the U-Net, such as SegNet and DeepLab v3+. Experimental results show that the proposed framework gives better and more reliable error analysis compared to conventional MSE evaluation. The experimental results indicated that the UNet + BBRM framework proposed in this study achieved an MSE of 0.0108, an accuracy of 98.92%, a Dice coefficient of 0.9369, and an IoU of 0.8831 in the segmentation task, respectively, compared with other methods. For the comparison with the local dataset, BBRM reduced the MSE from 0.022 to 0.015 and Dice (IoU) from 0.887 to 0.911 and 0.812 to 0.845. These findings underline the need for distribution-based error analysis and spatial alignment of segmentation methods in medical imaging applications.

Author 1: Dannial Asyraf Shahrul Anuar

Author 2: Nabilah Ibrahim

Author 3: Audrey Huong

Keywords: Kidney; BBRM; segmentation; U-Net; ultrasound imaging

Paper 24: Learning Mathematics Through Play: Design and Development of a Serious Game for Undergraduate Students

Abstract: Learning mathematics requires learners to have a good understanding of the concepts and problem-solving skills. However, mathematics is often associated with complexity and rarely perceived as connected to real-world applications. Educational games have the potential to make learning more enjoyable, motivating learners by offering challenges and control over their learning experience. Despite the growing research on mathematical games, existing educational games dominantly target primary and secondary levels, with limited focus on undergraduate calculus topics such as differentiation and integration. Visualizing mathematical concepts such as differentiation and integration through real-life scenarios and providing space for learners to learn while playing can be a suitable practice to persuade learners to learn mathematics. The aim of this study is to design and develop a game prototype that covers the topic of differentiation and integration. Method of instructional design: Analysis, Design, Development, Implementation, and Evaluation (ADDIE) is used to develop the game prototype, focusing on the systematic construction and expert validation. The serious game called CalcQuest Adventure was developed and distributed to eight experts to evaluate its suitability as a tool for learning the mathematical concepts of differentiation and integration. A verified instrument called the Suitability Evaluation Questionnaire (SEQ) is adapted in the study to evaluate CalcQuest Adventure. The result shows that the overall mean score is 3.85, which indicates positive feedback from experts. The findings contribute to the practical design insights for educators and developers by illustrating how real-life problems can be visualized and embedded into serious games for undergraduate calculus learning.

Author 1: Nur Nabilah Abdul Razak

Author 2: Nur Nabila Farhana Mohd Noh

Author 3: Nur Dayana Huda Mohd Zol Azhari

Author 4: Ratna Zuarni Ramli

Author 5: Ahmed M S Elaklouk

Author 6: Azyan Yusra Kapi

Keywords: Mathematics education; game design; differentiation and integration; calculus; STEM

Paper 25: Applying the AuRa Consensus Model for Digital Certificate Management in a Private Ethereum Blockchain

Abstract: The issue of fake certificates has been widely identified, and their prevalence has increased significantly in recent years. This growing trend has become a global concern due to its adverse impact on educational standards. A key factor contributing to the problem is the continued reliance on manual processes for issuing and verifying certificates. To address these challenges, this study proposes the use of an authority round (AuRa) consensus algorithm for managing certificate data on the Ethereum blockchain. AuRa, a member of the proof of authority (PoA) family, facilitates consensus among nodes distributed across multiple servers and networks. This mechanism plays a vital role in preserving the integrity and decentralization of the blockchain while ensuring the security of transactional data. Furthermore, the study investigates how AuRa enables efficient certificate data transactions within a private Ethereum environment. It also evaluates the algorithm's performance in terms of transaction speed per second (TPS) and throughput per second (TGS), demonstrating its effectiveness for managing certificate transactions on a blockchain network. Then the TPS and TGS results substantiate the suitability of AuRa for digital certificate generation, evidenced by its stable and efficient performance within a controlled private server environment.

Author 1: Robiah Arifin

Author 2: Wan Aezwani Wan Abu Bakar

Author 3: Mustafa Man

Author 4: Mohamad Afendee Mohamed

Author 5: Evizal Abdul Kadir

Keywords: Blockchain technology; fake certificate prevention; Authority round (AuRa) algorithm; ethereum private network; Proof of Authority (PoA); certificate verification system

Paper 26: A Bibliometric Analysis of Internet of Things and Augmented Reality Applications in Agriculture: A Scopus-Based Review (2019–2026)

Abstract: The integration of Internet of Things (IoT) and Augmented Reality (AR) technologies has emerged as a transformative approach in modern agriculture, enabling precision farming, real-time monitoring, and immersive decision-support systems. This study presents a bibliometric analysis of research trends related to IoT and AR applications in agriculture using data retrieved from the Scopus database. A total of 72 publications were gathered from 2020 to 2026. The metadata is then analyzed using bibliometric techniques to identify publication trends, document types, research themes, and emerging directions. The results reveal a steady increase in publications, with the highest number of records (20 documents) in 2024 and a slight drop to a total of 12 documents in 2025. Conference papers represent the largest proportion of publications (20 documents), followed by book chapters (17 documents), and reviews with 13 documents, reflecting the changing and experimental nature of this research area. Key research themes include precision agriculture, smart farming systems, digital twins, sensor networks, and real-time visualization for crop monitoring. The findings highlight the increasing role of AR interfaces integrated with IoT sensor data for enhanced agricultural decision-making. This study provides an inclusive overview of research developments and identifies potential future directions, including AI-driven AR visualization, digital agriculture ecosystems, and sustainable farming technologies.

Author 1: Ruziana Mohamad Rasli

Author 2: Sobihatun Nur Abdul Salam

Keywords: Bibliometric analysis; Internet of Things; Augmented Reality; smart agriculture; precision farming; Scopus database

Paper 27: Evolutionary Allocation with Dynamic Guidance for Pre-Scheduled Timetables

Abstract: The University Course Timetabling Problem (UCTP) is a well-known combinatorial optimization challenge that involves allocating courses to timeslots and rooms while satisfying various institutional constraints. At other institutions, general courses (e.g., language subjects) are prioritised during timetable allocation due to their high enrolment numbers from multiple faculties. However, at the MARA University of Technology, Sarawak Branch, Mathematics and Statistics (MAT/STA) courses are shared across several programs, with timeslot availability limited by pre-scheduled major courses in each program. This study presents a tailored evolutionary algorithm for a real institutional scenario, incorporating dynamic local search and guided variation operators to improve feasibility and solution quality. A case study was conducted using real datasets from the Department of Mathematical Sciences at MARA University of Technology, Sarawak Branch. Benchmark datasets from ITC2002 and ITC2007 (Track 2) are employed for comparative evaluation against existing methods. The algorithm successfully produced a feasible timetable and outperformed manually prepared schedules in terms of soft-constraint penalties. Results indicate strong performance on real datasets, producing a high-quality, balanced timetable with reduced preparation time, particularly in handling repeating students, lecturers' availability, and limited timeslots. Nevertheless, lower performance on benchmark datasets suggests the need for hybridization or additional operators to handle larger, more complex problems. Overall, the results demonstrate the effectiveness of the proposed approach in real-world timetabling while maintaining acceptable quality on benchmark instances.

Author 1: Anniza Hamdan

Author 2: Sze San Nah

Author 3: Goh Say Leng

Author 4: Emily Sing Kiang Siew

Keywords: Timetabling problem; evolutionary algorithm; dynamic; guided

Paper 28: Multi-Class Stress Detection Using Electrodermal Activity: Evaluation of A Hybrid Model Using UBFC-Phys Dataset

Abstract: Stress is a psychological and physiological response to internal or external pressures or challenges that exceed an individual's ability to cope. In response to these conditions, the human body produces physiological signals that reflect its internal states, often without conscious awareness. Among these signals, electrodermal activity (EDA) is known for its sensitivity to changes in stress levels. Traditional methods for assessing stress, such as questionnaires and self-reports, remain widely used, but they are often influenced by subjectivity and recall bias. This has led to increased interest in objective, data-driven approaches. This research proposes a hybrid stress detection model. The EDA signals were used from the UBFC-Phys dataset, which comprises data from 56 participants under three conditions: rest, moderate stress, and high stress. The signals were preprocessed using filtering, smoothing, and subject-level normalization to reduce inter-individual variability. A set of features was extracted, such as tonic component features. Multiple experiments were conducted on baseline, deep learning, and hybrid models not only for stress detection classification purposes but also for performance evaluation purposes. The proposed model achieved an accuracy of 92.16% in the multi-class classification task. The findings of this research contribute to mental health and well-being by providing an accurate model for stress detection that can be adapted to healthcare, education, and workplace settings, ensuring a healthier and more sustainable future.

Author 1: Kawther Alsayed

Author 2: Hamza Ghandorh

Author 3: Wael M.S. Yafooz

Keywords: Stress detection; physiological signal; Electrodermal Activity (EDA); machine learning; ensemble; deep learning; hybrid model

Paper 29: ATC Automation Based on Dynamic Aircraft Separation Techniques Accounting for Weather Conditions

Abstract: Air Traffic Management is becoming increasingly complex due to the continuous growth in air traffic demand and the variability of weather conditions. Traditional aircraft separation standards rely on fixed minimum distances that do not dynamically adapt to environmental factors. This study proposes a novel framework for dynamic aircraft separation that integrates real-time weather data into Air Traffic Control (ATC) automation systems. The proposed model adjusts separation distances based on meteorological parameters such as wind speed, turbulence intensity, and visibility. A multi-objective optimization approach is developed to minimize conflict risk while maximizing airspace capacity. Simulation results demonstrate that the proposed framework improves safety and operational efficiency compared to conventional methods. This study contributes to the advancement of intelligent air traffic management systems.

Author 1: Omayma Raziq

Author 2: Mohamed El Khaili

Author 3: Hasna Nhaila

Author 4: Azeddine Khiat

Keywords: Aircraft separation; air traffic control; weather impact; optimization; ATC automation; safety

Paper 30: A Computational Analysis of Housing Affordability: Robust Bootstrap Regression Modeling of the Safety Premium

Abstract: This study presents a methodologically rigorous and in-depth analysis of the socio-institutional determinants of housing affordability in nine EU Member States and Georgia over the period 2013-2023. The study transcends conventional macroeconomic models to identify and quantify the impact of non-financial indicators on the Price-to-Income ratio as the dependent variable. The study uses bootstrap regression with 1000 iterations, ensuring the stability of the coefficients and responding to the difficulties of data distribution in both transition and developed economies (unbalanced panel data, sample size of 89). Empirical results demonstrate a high explanatory power of the model (R-squared value of 0.670), where the Safety Index was identified as the main and most important driving factor of value formation (beta coefficient of 0.312, p-value of 0.007). The conclusion that institutional safety functions as a "nonfinancial capital" that exhibits a strong predictive association with the volatility of real estate prices is significant. The research findings also indicate a paradigm shift in real estate valuation, validating the "safety premium" and supporting the Hedonic Pricing Theory. This study argues that institutional stability in the housing ecosystems of transition countries outweighs the importance of traditional financial liquidity as a price determinant.

Author 1: Natia Terterashvili

Author 2: Shota Shaburishvili

Keywords: Housing affordability; robust bootstrap regression; safety index; EU real estate market; Georgia; value driver

Paper 31: An Enhanced Kalman Filter-Based Hybrid Battery Management System for Energy Optimization in IoT-Enabled Smart Agriculture

Abstract: Battery-powered sensor nodes operating in a remote environment are used in IoT-based smart agriculture deployments, where the maintenance of these devices is impractical. Thus, precise SOC (state-of-charge) estimation is critical to enhance both energy efficiency and system stability. A new hybrid BMS for SOC estimation based on a transaction-oriented adaptive Kalman Filter that merges traditional voltage-based SOC estimation has been developed in this study. Results show that our proposed method removes noise sensitivity and reduces SOC drift subjected to dynamic IoT workloads by integrating model prediction with measurement correction. Application experiments based on 30 days' real sensor workload data indicate that the cumulative battery consumption index of the proposed algorithm decreased from 78% to 56%, and provided a performance improvement of approximately 22.7% compared to the baseline method. Moreover, the proposed method shows better stability in terms of daily consumption fluctuation and drift. The findings demonstrate that the improved Kalman Filter–based hybrid BMS serves as an efficient and effective alternative for achieving long-term energy optimization in IoT-enabled smart agriculture applications.

Author 1: Mohd Kamir Yusof

Author 2: Nur Yasmin Salleh

Author 3: Senny Luckyardi

Author 4: Wan Mohd Amir Fazamin Wan Hamzah

Author 5: Mustafa Man

Keywords: Internet of Things; smart agriculture; battery management system; state-of-charge estimation; Kalman Filter; energy optimization

Paper 32: Enhanced Real-Time Fire Detection Systems Using Deep Learning and Differentiating Between Dangerous and Non-Dangerous Fires

Abstract: Fire is a major hazard in many disasters, creating risks to public safety and the surrounding environment. This study aims to improve the accuracy and reliability of fire detection using deep learning, addressing the key limitations of traditional sensor systems, such as latency and poor adaptability. The proposed model presents a customized Fire-Smoke-YOLOv8x Model trained from scratch on a dataset of 100,000 images representing diverse fire and smoke conditions paired with adaptive algorithms to differentiate dangerous from non-dangerous fire scenarios and achieved a performance with 98.2 % precision, 97.3% recall, and 94% mAP@50. The model processes RGB video under varied lighting and environments and achieves strong detection results across a wide range of fire and smoke scenarios. Incorporating thermal or multispectral data, as noted in future work, could further improve performance under extremely low visibility conditions. This framework supports real-time surveillance in smart cities, transport hubs, and industrial safety, where fast and accurate detection is critical. It combines fine-grained detection with risk-aware post-processing, providing a high-performance and scalable solution for real-world fire detection.

Author 1: Mohamed Youssef

Author 2: Mohamed Marie

Author 3: Sarah Naiem

Keywords: Deep learning; fire detection; smoke detection; computer vision; real-time detection; attention mechanism; fire risk classification

Paper 33: A Novel Hybrid Deep Learning Validation Model for Real-Time and Synthetic Image Inputs in Capsicum Plant Disease Diagnosis

Abstract: The diagnosis of plant diseases in Capsicum species remains a critical challenge in precision agriculture due to variability in environmental conditions and limited availability of high-quality datasets. The traditional convolutional neural network (CNN) has been demonstrated to have satisfactory performance, but it cannot capture robust performance in both real-time and synthetic images. This study introduces a novel hybrid deep learning validation model based on Convolutional neural networks (CNN) combined with Capsule networks (CapsNet) and Vision Transformer (ViT) backbone. The framework is intended to be able to validate multi-source image inputs and improve the reliability of the classification in natural and synthetic image environments. In contrast to previous plant disease detection models, which can only be trained on real-time plant images, the proposed CNN–ViT–CapsNet framework features a dual-domain validation process that is able to classify both real-time and the GAN-generated synthetic Capsicum leaf images. CNN-based local feature extraction, ViT-based global contextual learning, and CapsNet-based spatial validation ensure robustness against illumination, orientation, and background conditions.

Author 1: Prashant Vikhe

Author 2: Baisa Gunjal

Keywords: Capsicum plant disease; hybrid deep learning; real-time validation; synthetic images; image processing; CNN-CapsNet-ViT

Paper 34: Large-Scale Static Malware Detection Using Classical Machine Learning Models: An Evaluation on the EMBER Dataset

Abstract: Malware detection is a major difficulty in cybersecurity as malicious software continues to evolve in scale, diversity, and sophistication. While deep learning and highly complex architectures are becoming increasingly important in recent work, the practical efficiency of conventional machine learning methods for large-scale static malware detection remains underexplored. We perform a comparative evaluation of four machine learning models (Random Forest, XGBoost, Logistic Regression, and Decision Tree) on approximately 600,000 Portable Executable (PE) samples from the EMBER dataset. To enable a fair comparison of the models, we created a common experiment setup including standardised preprocessing, repeated evaluation with numerous random seeds, selective hyperparameter optimisation, feature importance analysis, and confusion matrix-based error analysis. The experimental results show a strong benefit of ensemble-based approaches for the structured feature representation provided by the EMBER dataset. Random Forest showed the best overall performance, with 96.74 % accuracy, 96.71 % F1-score, and a ROC-AUC of 0.9953, retaining a very steady behavior in repeated runs. XGBoost likewise demonstrated good predictive capacity with less training time but did not outperform Random Forest even with careful hyperparameter adjustment. On the other hand, Logistic Regression performed significantly worse, suggesting that linear decision boundaries were insufficient to capture the deep structural relationships encoded in static malware traits. Further study of the confusion matrix shows a balanced classification behavior with relatively low false negative rates, which is significant for operational malware detection situations. The feature importance analysis suggested that entropy-based features, PE structure metadata, and import-based features played an important role in the malware classification judgments. In conclusion, the results suggest that well-designed classical ensemble approaches are still quite competitive for scalable and interpretable static malware detection even with the rising usage of more and more powerful machine learning architectures.

Author 1: Achmad Fauzan

Author 2: Tito Pinandita

Author 3: Aulia Desy Nur Utomo

Keywords: Malware detection; static analysis; EMBER benchmark; ensemble learning; PE file analysis; random forest; XGBoost

Paper 35: A Data-Driven Visual Analytics Framework for Transaction-Level Retail Profit Modeling and Decision Support

Abstract: Retail companies are becoming increasingly dependent on data science to inform their pricing, assortment, and regional strategy decisions. Profitability drivers, however, are often difficult to interpret because they span product-level, geographic, discount, and operational environments. In most real-life applications, analytics and visualization are treated as two distinct processes, thereby restricting interpretability and undermining their ability to support decision-making. This study provides a decision-level visual analytics model of retail profit analysis at the transaction level. The model is illustrated on the publicly available Superstore dataset as a benchmark, which contains 9,994 order-line records between 2014 and 2017 with the variables time, product, geographic, customer-segment, shipping, sales, discount, and profit. The workflow combines feature engineering, hierarchical slicing, variance-based comparison, and five publication-ready dashboards covering temporal trends, product profitability, geographic heterogeneity, discount sensitivity, and fulfillment context. The overall profit margin in the analyzed dataset is 12.47%, and 18.72% of order lines are loss-making. The findings show uneven profitability across product groups and states, with loss pockets observed in sub-categories such as Tables and Bookcases, as well as in states like Texas and Ohio. There is also a distinct nonlinear discount behavior: profitability tends to be positive at discount rates below about 20%, then margins decline sharply, and loss rates increase exponentially at higher discount rates. These benchmark-specific results illustrate how integrated visual analytics can support structured inspection of pricing, portfolio, and region-related patterns within transaction-level retail data.

Author 1: Donia Badawood

Keywords: Visual analytics; retail profitability; discount sensitivity; decision support systems; transaction-level analytics; retail data science; profit diagnostics; interactive dashboards

Paper 36: Fostering Programming Interest Through Block Coding in Young Learners

Abstract: Coding has become a fundamental skill for cultivating computational thinking in young learners. This study aims to examine the impact of coding education on programming interest among fifth-grade students in Mongolia with no prior coding experience. A 16-week pilot course using block-based coding activities on Code.org involved 195 5th-grade students across three schools. A pre-test/post-test design compared an experimental group (Nexperimental=95) with a control group (Ncontroll=100). Results indicate significant improvements in coding skills (Cohen’s d = 0.84) and computational thinking, with average programming skill ratings increasing from 5.19 to 7.07 out of 10. Qualitative data show heightened engagement, perceived creativity, and motivation, with students describing coding as gamified activities alongside an increase in self-rated programming interest. Moreover, post-intervention interviews reveal emerging awareness of gender inclusivity in programming. These findings suggest that structured block-based coding can effectively enhance computational thinking and interest in programming among young learners, providing evidence to support early integration of programming in primary education curricula.

Author 1: Dorjpalam Tserendejid

Author 2: Khajidmaa Otgonbaatar

Author 3: Magsarjav Bataa

Keywords: Computational thinking; problem-solving; block-based coding; coding education; programming interest

Paper 37: MYPO-Net: A Robust Deep Learning Approach for Multi-Yoga Pose Detection and Occlusion Handling

Abstract: Yoga has become a well-known holistic process worldwide and has been appreciated due to its physical, psychological, and injury-preventive effects. The swift development of online fitness applications has created a growing need for automated systems that can precisely identify and analyze yoga poses. The current methods, however, are limited in terms of the limited diversity of datasets, insufficient occlusion, low performance in multi-person settings, and a lack of feedback mechanisms to offer corrective feedback. In order to overcome these shortcomings, this study introduces MYPO-Net, an artificial intelligence (AI) based deep learning model that uses the efficiency of MobileNet and the classification performance of EfficientNetB0. The model is trained and tested on the Yoga82 data, with a detailed preprocessing pipeline, such as resizing, normalization, and data augmentation, to improve resilience to real-world variations. Experimental evidence shows a classification accuracy of 97.65, which is higher than a variety of baseline architectures (VGG16: 87%, InceptionV3: 82%, ResNet50: 58) and has high computational efficiency. Confusion matrix analysis shows that there is valid detection in 16 yoga position classes. The persisting issues in real-time implementation and poor image quality settings are distinguished as future work directions. MYPO-Net is a highly scalable, affordable, and open-source platform to support digital yoga teaching, fitness apps, and rehabilitative health care.

Author 1: Rehana Danial

Author 2: Nosheen Qamar

Author 3: Nosheen Sabahat

Author 4: Faria Nazir

Author 5: Ali Salem Bin Sama

Author 6: Lamia Hassan Rahamatalla

Author 7: Osman Elwasila

Author 8: Abdulrahman Alojail

Author 9: Marwan Abu-Zanona

Keywords: Yoga pose classification; deep learning; mobilenet; efficientnetb0; occlusion handling; transfer learning; convolutional neural networks; human pose estimation

Paper 38: Confirmatory Factor Analysis of Phishing Susceptibility in Indonesia

Abstract: This study analyzes phishing susceptibility among Indonesian internet users through Confirmatory Factor Analysis (CFA) and covariance-based Structural Equation Modeling (SEM). The latent factors examined include perceived severity of the threat (PST), perceived barriers (PBR), perceived benefits (PBN), self-efficacy (SE), past success in detection (PSD), and phishing desensitization (PD), with phishing susceptibility (PS) as the dependent variable, derived from survey data of 150 respondents who had encountered phishing attacks. CFA results indicate a good model fit (RMSEA=0.064, CFI=0.92, etc.), while tests of six hypotheses reveal no significant positive correlations between these factors and PS. These findings challenge prior literature assumptions and underscore the need for mediating factors such as digital literacy or cultural norms to mitigate phishing vulnerability in Indonesia, offering implications for internet service providers and government policy.

Author 1: Harris Simaremare

Author 2: Muhammad Fikri

Author 3: Zarina Shukur

Keywords: Confirmatory Factor Analysis; cybersecurity awareness; Structural Equation Modeling; phishing susceptibility

Paper 39: AgricultureEarlyWarning: A Web-Based Climate Advisory and Early Warning Platform for Albanian Farmers

Abstract: This study reports the design and implementation of AgricultureEarlyWarning, a web-based prototype developed to operationalize climate information services and early warning logic for farmer-centered agricultural risk management in Albania. The system integrates parcel-level registration and geolocation, daily and hourly forecast ingestion, crop-stage-sensitive risk evaluation, satellite-derived indicators, dashboard analytics, AI-assisted agronomic explanation, vulnerability profiling, escalation support, and scheduled alert dissemination within a single application. The implementation stack combines ASP.NET MVC on .NET Framework 4.8, SQL Server persistence, Hangfire background scheduling, Open-Meteo weather services, Sentinel Hub APIs, and OpenStreetMap reverse geocoding. Rather than proposing a novel forecasting algorithm or claiming a complete national early warning system, the study contributes an implementation-oriented platform that links data acquisition, contextual risk interpretation, farmer-facing visualization, and asynchronous communication. The findings indicate that the prototype supports localized monitoring, advisory generation, alert scheduling, and analytics, while field validation, long-term adoption assessment, and full institutional response governance remain priorities for future work.

Author 1: Irida Gjermeni

Keywords: AgriculturalEarlyWarning system; climate information services; farmer advisory; remote sensing; ASP.NET MVC; Albania; decision support

Paper 40: Spatial Anemia and Osmotic Fragility of Erythrocytes in Microgravity Conditions: A Systematic Review

Abstract: The study of the effects of microgravity on the human body has become increasingly relevant in biomedical research, particularly with regard to hematological changes. Therefore, the objective was to analyze the available scientific evidence on the impact of microgravity, both real and simulated, on erythrocytes with emphasis on osmotic fragility, cell membrane stability, and hemolysis. Likewise, 40 articles were selected using the PRISMA methodology, using the Scopus, PubMed, Web of Science, and IEEE Xplore databases, applying inclusion and exclusion criteria that had been previously established. Tools such as VOSviewer were also used, which facilitated the visualization of patterns of collaboration and scientific production. The results showed that microgravity produces remarkable changes in erythrocytes, such as an increase in hemolysis, alterations in the capacity for cell deformation, increased oxidative stress, and variations in the membrane. However, most research shows indirect information when it comes to osmotic fragility, with few applications of standardized measurements such as OFT or MCF-H50. In conclusion, severity influences the integrity and functionality of erythrocytes, leading to spatial anemia, although studies with greater standardization in their methodology are needed.

Author 1: Lilian Ocares-Cunyarachi

Author 2: Natalia Vargas-Cuentas

Author 3: Avid Roman-Gonzales

Author 4: Georgina Chávez

Author 5: Brigitte Torres-Pinedo

Author 6: Ana Huamani-Huaracca

Keywords: Erythrocytes; hemolysis; microgravity; osmotic fragility; spatial anemia

Paper 41: Artificial Intelligence-Assisted Community Needs Assessment and Extension Planning: Evidence from Wesleyan University-Philippines Partner Communities

Abstract: This study conducted an Artificial Intelligence-assisted community needs assessment of partner-community respondents of Wesleyan University-Philippines, with emphasis on the Tricycle Operators and Drivers Association. Using a descriptive-quantitative design, five objectives were addressed: to describe the socioeconomic profile of the respondents; to determine livelihood and income conditions; to identify health, educational, and public-service needs; to rank priority community needs; and to develop a transparent, data-driven Artificial Intelligence-assisted framework. A cleaned dataset of 151 respondents was analyzed, of whom 46 were formally affiliated with tricycle operator and driver associations. Frequency, percentage, mean, and a proportion-based priority score summarized the data, while an Artificial Intelligence-assisted workflow supported response coding, respondent clustering, pattern detection, and need-to-intervention matching. Results showed that the transport subgroup consisted entirely of male drivers with lower average income, stronger income-expenditure pressure, greater reliance on borrowing and relatives, and recurring health concerns, including respiratory illness, hypertension, dental problems, and diabetes. Educational needs centered on books and learning materials. Ranked priorities were transport-oriented support, public-service linkage, health and wellness, educational assistance, livelihood and financial stability, skills training, and environmental support. The study concludes that pairing descriptive statistics with a reproducible Artificial Intelligence-assisted procedure produces more targeted, equitable, and responsive extension planning, and offers a practical template for other higher-education institutions.

Author 1: Eufemia Ayro

Author 2: Karl Leugim Bernarte

Author 3: Hazel May Babiera

Author 4: Evangeline Agpoon

Author 5: Jhon Carlo Villa

Author 6: Maureen Bondoc

Author 7: Jennyfer Villalon

Author 8: Jose Arsenio Adriano

Author 9: Christian Navarro

Keywords: Artificial Intelligence; Community Needs Assessment; data-driven extension program; tricycle drivers; needs prioritization; decision support

Paper 42: A Comparative Analysis of Decision Tree, Random Forest, and Logistic Regression Models in Predicting Business Readiness for Digital Technology Integration

Abstract: This study compared the performance of Decision Tree, Random Forest, and Logistic Regression models in predicting business readiness for digital technology integration using survey data from 400 business respondents in Pangasinan, Philippines. The analysis utilized variables related to technology utilization, perceived helpfulness, willingness to integrate technology, and challenges encountered in adopting digital marketing, e-commerce, and digital payment technologies. Business readiness was operationalized from respondents’ willingness scores and classified into Ready and Less Ready categories. Descriptive results revealed very low technology utilization despite high perceived helpfulness, indicating a gap between awareness of digital technologies and their actual adoption. Lack of awareness emerged as the most frequently reported barrier, followed by data security concerns. Using an 80:20 train test split, the machine learning models achieved moderate predictive performance, with Decision Tree and Random Forest attaining the highest accuracy of 63.75%. Random Forest produced the best overall performance, achieving the highest weighted F1 score and demonstrating a more balanced classification capability than the other models. The findings highlight the potential of machine learning as a decision support tool for assessing business readiness and generating evidence-based insights that can support digital transformation planning, technology adoption strategies, and capacity building initiatives among businesses.

Author 1: Gloria M. Ducut

Keywords: Business readiness; digital technology integration; machine learning

Paper 43: Canthus-Scaled 468-Landmark FaceMesh Framework for Pupillary Distance Estimation Using Nested AutoML Calibration

Abstract: Pupillary distance (PD) is an important ocular measurement for optical dispensing and vision-related applications, but standard MediaPipe FaceMesh outputs do not provide true pupil-centre or iris-boundary landmarks when only the 468-landmark representation is available. This study proposes a canthus-scaled 468-landmark framework for estimating PD using facial landmarks and Malay young-adult normative palpebral fissure width. A dataset of 44 subjects was used, where each record contained ground-truth PD and 1,404 coordinate values representing 468 FaceMesh landmarks in three dimensions. Since true pupil centres were unavailable, medial and lateral canthus landmarks were used to construct eye-centre proxies and to compute a subject-specific millimetre scale. A direct canthus-scaled proxy was first evaluated as a deterministic baseline, after which canthus-scaled geometric features were used in a nested AutoML calibration framework. Model development used repeated nested cross-validation, with an outer repeated 5-fold design and an inner 4-fold model-selection loop. The direct proxy achieved a mean absolute error (MAE) of 4.26 mm and showed systematic overestimation. The calibrated nested AutoML model improved performance, achieving a subject-level MAE of 3.510 mm, root mean squared error of 4.22 mm, a bias of −0.08 mm, and 75.0% of predictions within ±5 mm. The calibrated nested AutoML model improved overall error and reduced systematic bias compared with the direct canthus-scaled proxy. However, the Bland–Altman limits of agreement remained wide, indicating that the proposed method should be interpreted as an approximate proxy-based estimation approach rather than a substitute for clinical pupillometer- or ruler-based PD measurement. The framework is most relevant for research settings or datasets where only standard 468-landmark FaceMesh data are available, and iris-refined landmarks are absent.

Author 1: Mohd Izzuddin Mohd Tamrin

Author 2: Sherzod Turaev

Author 3: Takumi Sase

Author 4: Mohd Zulfaezal Che Azemin

Author 5: Tengku Mohd Tengku Sembok

Keywords: Pupillary distance; MediaPipe FaceMesh; facial landmarks; canthus scaling; palpebral fissure width; AutoML; nested cross-validation; computer vision

Paper 44: A Hybrid CNN-Ensemble Framework for Robust DeepFake Image Detection

Abstract: The fast development of deepfake technologies has caused growing concerns related to the authentication of digital media, the integrity of personal identification, and the spreading of disinformation. Therefore, there is a growing need for effective automatic detectors of deepfakes. Meanwhile, existing techniques of deepfake detection encounter a large number of difficulties. First, it is difficult to distinguish the subtle manipulation details of faces. Another problem is poor generalization when applying models to novel datasets or highly realistic, generated synthetic faces. Another issue is low accuracy in the case of imbalanced data, where samples of one class dominate others. Thus, to address those problems, this study proposes a novel hybrid approach based on a combination of deep learning (DL) and conventional machine learning (ML) for detecting deepfake images. More precisely, two pre-trained CNNs (MobileNetV2 and ResNet50) were applied to generate features and classify them via the Random Forest (RF) algorithm. The experiments have been conducted on a benchmark of 6,557 facial images marked as either real or fake. The findings reveal that the MobileNetV2+RF achieved the highest accuracy 99%, followed by MobileNetV2 with 98% and ResNet50 with 97% accuracy. This suggests that the hybrid architecture helps to increase the effectiveness of the solution. A statistical significance test reveals that none of the models' performances differ significantly from each other (p > 0.05). Overall, the proposed system demonstrates excellent metrics concerning accuracy and precision.

Author 1: Mohammad Alsulami

Keywords: Deep learning models; machine learning classifiers; image processing; data augmentation

Paper 45: A Bibliometric Mapping of Transformer-Based Misinformation Detection: Trends and Gaps

Abstract: The rapid spread of misinformation on social platforms has intensified research on automated detection, with Transformer-based architectures becoming a primary technical foundation. However, the volume and diversity of publications make it difficult to track the field’s evolution. It is also challenging to identify the contributors, influential studies, and remaining gaps that shape its knowledge base. This study presents a bibliometric mapping of Transformer-based research on misinformation detection, based on Scopus records retrieved on 3rd March 2026 (n = 3,637). Performance indicators and science mapping analysis are used to profile publication growth, subject-area distribution, leading countries, institutions, and authors, citation impact, and thematic structure. Results show that conference proceedings (47.18%) and journals (41.93%) dominate dissemination, reflecting both rapid-cycle publishing and archival consolidation. Computer Science accounts for most publications (84.99%), while Social Sciences and Medicine contribute non-trivial shares, indicating applied and societal engagement. China, the United States, and India lead in national output, and institutional productivity is concentrated in a small set of research hubs. Citation indicators (h-index = 82; g-index = 154) suggest a broad influence with a distinct, highly cited core. Keyword mapping confirms a stable emphasis on NLP-driven detection in social media and shows recent growth in multimodal approaches and generative-AI-related topics. The findings provide an evidence-based overview of current trends and emerging directions, helping researchers position future work and prioritize underexplored problems.

Author 1: Borhan Ab Rahman

Author 2: Mohd Zakree Ahmad Nazri

Author 3: Mohd Ridzwan Yaakub

Keywords: Bibliometric analysis; misinformation detection; Transformer architectures; keyword co-occurrence; citation analysis; Scopus

Paper 46: Web System to Optimize Information Management for Beca 18 Selection-Stage Applicants, 2025

Abstract: The objective of this study was to determine whether the implementation of a web system optimizes information management by reducing the time required to access, verify, and interpret selection-stage information for applicants of the Beca 18 2025 scholarship call. The research followed an applied, quantitative, pre-experimental design with pre-test and post-test measurements in a single group. The sample comprised 361 preselected applicants from the ordinary modality in the Lima region, selected through simple random sampling. A direct observation sheet recorded the time, in minutes, required by each participant to complete 15 tasks associated with information accessibility, information currency, and information quality before and after using the web system. The mean information-management time decreased from 130.40 to 50.46 minutes, with a statistically significant mean reduction of 79.94 minutes. The three evaluated dimensions also showed substantial reductions: information accessibility decreased by 70.3 per cent, information currency by 80.2 per cent, and information quality by 52.2 per cent. The findings indicate that, within the scope of time-based task performance, the web system optimized the efficiency with which applicants accessed, verified, and interpreted selection-stage information.

Author 1: Jack Quispe-Vivas

Author 2: Cristian Sedano-Contreras

Author 3: Junior Segovia-Chalco

Author 4: Brian Meneses-Claudio

Keywords: Web system; information management; Beca 18; information accessibility; information currency; information quality; pre-experimental design

Paper 47: A Modified Lyapunov-Based MEC Offloading Algorithm for Severity-Aware QoS in 5G/B5G IoT Systems

Abstract: The rapid growth of latency-sensitive and computation-intensive IoT applications in 5G and Beyond-5G (B5G) networks has increased the demand for efficient Multi-access Edge Computing (MEC) offloading strategies. Current MEC frameworks have several limitations: 1) binary QoS modeling without considering deadline violation severity, 2) a lack of severity-aware optimization in IoT applications, 3) insufficient consideration of different task criticality, and 4) poor handling of dynamic latency and energy trade-off in large-scale IoT environments. This study proposes a modified Lyapunov-based severity-aware MEC offloading framework for heterogeneous 5G/B5G IoT systems. The proposed framework utilizes task deadlines, task criticality, queue states, wireless channel conditions, and MEC resource availability as input for adaptive offloading optimization. A QoS Violation Severity Index is introduced to jointly capture deadline violation magnitude and task criticality. Furthermore, severity-aware virtual queues are integrated with a modified Lyapunov Drift-Plus-Penalty optimization framework to dynamically minimize QoS violation severity while balancing latency and energy consumption. Experimental evaluation demonstrates that the proposed framework significantly reduces average task delay to 82 ms, energy consumption to 6.0 mJ, and QoS violation rate to 4.8%, while improving long-term system stability compared with existing MEC offloading approaches in dynamic 5G/B5G IoT environments.

Author 1: Sahana S Reddy

Author 2: R. Sukumar

Keywords: 5G and Beyond-5G; multi-access edge computing; quality of service; task offloading; severity estimation; lyapunov optimization

Paper 48: Bridging Topic Modelling Outputs to Bayesian Hierarchical Model Using LLM and WordNet Parameter Labelling

Abstract: This study investigates the challenge of generating accurate and interpretable topic labels for integration into Bayesian Hierarchical Models (BHM), a critical step for interpretable probabilistic risk modelling from unstructured textual data. Using a corpus of 35,667 Malaysian business news articles published between 2019 and 2023, four topic modelling approaches, such as Latent Dirichlet Allocation (LDA), Non-negative Matrix Factorization (NMF), Top2Vec, and BERTopic, were evaluated. Among these, NMF produced the most coherent and thematically consistent topics. To address the topic-labelling challenge, this study proposes an NLP-BHM framework that maps topic model outputs into a hierarchical Bayesian structure comprising interpretable topic labels and higher-level risk categories. Two semantic labelling strategies were examined: Large Language Model (LLM) prompting and WordNet-based semantic analysis. The proposed approaches enabled systematic topic interpretation and semantic clustering within the BHM framework. A case study on Malaysian business risks demonstrates that LLM-based labelling produced more coherent and contextually relevant results, while WordNet-based labelling provided a semantically consistent but vocabulary-limited alternative. Comparative results based on Mean Opinion Scores (MOS) highlight the effectiveness of LLM-based semantic labelling in improving interpretability for probabilistic business risk analysis.

Author 1: Vadrianey Asas

Author 2: Sarah Samson Juan

Author 3: Stephanie Chua

Author 4: Evan Lau

Author 5: Jane Labadin

Keywords: Bayesian hierarchical model; natural language processing; topic modelling; large language model; WordNet

Paper 49: Artificial Intelligence and the Transformation of Academic Integrity in Higher Education: A Systematic Review

Abstract: This study examines how Artificial Intelligence (AI) is transforming academic integrity in higher education, altering both learning opportunities and the risks associated with misconduct. As creative AI tools become embedded in everyday academic work, they provide valuable support for writing, research assistance, and skills development. Still, they also challenge long-held assumptions about authorship, originality, and assessment. Emerging evidence suggests that students are using AI in a variety of ways, from supporting legitimate learning to producing fully automated assignments. However, AI-driven integrity technologies, such as plagiarism detectors and authorship checking models, are becoming more effective but continue to face issues of bias, false positives, and limited transparency. This rapid shift has created a gap between technological change and academic readiness, highlighting the need for institutions to rethink assessment design, improve integrity frameworks, and foster a culture of responsible AI use, rather than relying solely on surveillance and sanctions. This review compiles the latest studies published between 2020 and 2025 to map current practices, risks, and policy responses. The findings suggest that academic integrity in the age of artificial intelligence (AI) cannot be focused solely on preventing fraud. But this needs to expand to support ethical digital literacy, redesign learning tasks that require human reasoning, and ensure fairness in automated decision-making systems. The study concludes with recommendations for educators, researchers, and policymakers to balance innovation with responsibility to ensure that AI becomes a tool for transforming learning, rather than a threat to academic values.

Author 1: Wannakorn Phornprasert

Author 2: Wongpanya S. Nuankaew

Author 3: Pratya Nuankaew

Keywords: AI in education; academic integrity; Generative AI; ethical digital literacy; assessment transformation

Paper 50: A Secure Integrated Cloud Storage Framework Using Lorenz Chaotic Key Generation, AES-256-GCM, and PBFT-Based Blockchain Verification

Abstract: Cloud storage security remains a major challenge owing to threats related to confidentiality, integrity, and unauthorized access. In this study, a novel secure cloud storage system is proposed by integrating Lorenz 3D chaotic key generation, AES-256-GCM authenticated encryption, and blockchain verification based on PBFT. High-entropy keys were generated using the Lorenz chaotic system and then passed through SHA-256 to create secure 256-bit AES keys. The data is encrypted using the AES-256-GCM algorithm and stored in the cloud, and integrity metadata is stored in a blockchain that offers tamper detection, auditability, and Byzantine fault tolerance. The experimental results demonstrate strong cryptographic performance with near-ideal entropy (7.99), a high avalanche effect (50.05%), successful NIST randomness validation, and low blockchain overhead. The results of the security analysis prove resistance against brute force, statistical, differential, replay, known plaintext, chosen ciphertext, and Byzantine attacks. The proposed framework offers a secure, efficient, and reliable multilayered security solution for cloud storage systems.

Author 1: Walde Rajesh Baliram

Author 2: Bashir Alam

Author 3: Mohammad Najmud Doja

Keywords: Cloud security; AES-256-GCM; Lorenz chaotic system; blockchain; PBFT; data integrity; secure cloud storage; cryptography

Paper 51: Comparative Evaluation of Traditional and Transformer-Based Models for Risk-Level Classification of Uzbek Telegram Messages

Abstract: The rapid growth of Telegram-based communication has increased the dissemination of harmful and risky content, particularly in low-resource languages such as Uzbek. This study investigates the automatic classification of Uzbek Telegram messages according to risk level using both traditional machine learning and transformer-based models. A dataset consisting of 10,000 real Telegram messages was collected and manually annotated into two classes: Safe and Dangerous. To improve data quality and consistency, preprocessing techniques including URL removal, emoji normalization, stop-word filtering, and script unification were applied. The study compares the performance of TF-IDF + Logistic Regression, FastText, mBERT, and XLM-RoBERTa for harmful content detection in Uzbek Telegram texts. Experimental results show that transformer-based models significantly outperform traditional approaches. Among all evaluated models, XLM-RoBERTa achieved the highest performance, with an Accuracy of 91.2%, a precision of 90.8%, a recall of 91.5%, and an F1-score of 91.1%, while mBERT achieved an Accuracy of 84.9% and an F1-score of 84.6%. The results demonstrate the effectiveness of contextual transformer architectures for identifying harmful content in low-resource language environments. The findings confirm that transformer-based models provide a reliable solution for automatic risk-level classification of Uzbek social media texts and can support practical applications in content moderation, information security, and social media monitoring systems.

Author 1: Feruzakhon A. Qoyliyeva

Author 2: Ozod J.Babomuradov

Author 3: Akmal A. Savurbayev

Keywords: Uzbek language; telegram messages; risk-level classification; harmful content detection; machine learning; transformer models; mBERT; XLM-RoBERTa

Paper 52: MedChain: A Privacy-Preserving Framework for Scalable Electronic Health Record Sharing on Blockchain and InterPlanetary File System

Abstract: The secure and scalable sharing of electronic health records (EHRs) remains a fundamental challenge in modern health-care systems due to conflicting requirements of privacy, regulatory compliance (HIPAA, GDPR), and real-time clinical access. Existing blockchain-based solutions suffer from three critical limitations, including static access control policies that cannot adapt to emergency scenarios, complete transparency of policy evaluation leading to attribute leakage, and linear degradation of throughput with policy complexity. The proposed MedChain, a framework introducing four novel contributions: Dynamic Attribute-Based Proxy Re-Encryption (DAB-PRE) enabling time-bound and emergency breakout access without re-encrypting underlying data; Zero-Knowledge Policy Verification (ZK-PV) allowing smart contracts to validate access requests without revealing sensitive attribute values; Adaptive EHR Sharding that dynamically adjusts IPFS shard sizes based on access frequency, reducing cold-start latency by 64%; and Layered Revocation Cascade providing granular revocation at patient, provider, and record-type levels within 2.3 seconds. Using the MIMIC-III benchmark (46,520 patients, 100,000 access traces), we demonstrate: 99.4% storage reduction vs. pure blockchain, 0.95s median latency (43% improve-ment over prior work), 2,450 requests/second throughput with 150 IPFS nodes, and provable security under the q-DBDH assumption. Comprehensive comparisons with MedRec, FHIRChain, Ancile, and PriHac confirm MedChain’s superiority across all metrics.

Author 1: S. Venkateswaran

Author 2: N. Vijayaraj

Keywords: Blockchain; electronic health records; IPFS; dynamic attribute-based proxy re-encryption; zero-knowledge proofs; adaptive sharding; MIMIC-III

Paper 53: Token-Level PII Detection with Symbolic, Sequential, and Transformer-Based Ensemble Models

Abstract: The rapid increase in unstructured digital information has led to an urgent demand for effective systems for safeguarding Personally Identifiable Information (PII) across multiple sectors and application domains. Existing single-model approaches frequently fail to resolve entity-type ambiguity in unstructured text, particularly when a token's PII status is context-dependent rather than syntactically predictable. This study presents EnsemblePII, a weighted voting ensemble that combines rule-based patterns, dictionary matching, Conditional Random Fields, Bi-LSTM sequence models, and transformer-based token classification. The ensemble applies a class-specific weighted voting strategy in which each model's per-entity influence is proportional to its per-class F1-score on a held-out validation set. The approach is assessed by using the ai4privacy/pii-masking-43k data set and the Mendeley Synthetic Financial Documents data set. EnsemblePII achieves a weighted F1 of 0.9749 on the general-purpose HF test set, marginally exceeding the strongest individual component (DistilBERT, F1 = 0.9744) on the in-distribution evaluation, and outperforming a published entity-span-level hybrid baseline by over 6 percentage points on the multi-class token-level task. On the Mendeley financial test set, Bi-LSTM and DistilBERT achieve F1 scores of 0.9959 and 0.9703, respectively, while the ensemble records 0.8433, revealing calibration sensitivity to entity classes absent from the weight calibration corpus. The results indicate that ensemble-based PII detection can improve in-distribution performance, but stable cross-domain generalization requires domain-aware calibration of ensemble weights; the DistilBERT component achieves the highest average F1 (0.9724) across both test sets, underscoring the value of transformer-based models for cross-domain PII detection.

Author 1: Hessah Abdullah Alshamrani

Author 2: Mona Alnahari

Keywords: Personally identifiable information; PII detection; weighted voting ensemble; named entity recognition; BIO tagging; transformer models; data privacy

Paper 54: Holistic Model of Advanced Analytics to Optimize Organizational Decision-Making

Abstract: In increasingly data-intensive organizational environments, decision-making processes require analytical frameworks capable of integrating data governance, advanced analytics, and strategic interpretation under a unified structure. This study proposes a holistic advanced analytics model designed to optimize organizational decision-making through the integration of data quality, data integration, analytical capabilities, and data-driven storytelling within a continuous decision-support lifecycle. The proposed model was developed using the Design Science Research (DSR) methodology and structured according to the intelligence, design, and choice phases of the classical decision-making process. The framework incorporates internationally recognized standards and methodologies, including ISO/IEC 25012, ISO/IEC 11179, CRISP-DM, DataOps, and analytics value chain principles, enabling methodological interoperability and adaptive analytical governance. The resulting artifact was conceptually validated through expert judgment involving seven specialists in analytics, business intelligence, and organizational decision-making. The evaluation produced average scores ranging from 3.29 to 4.57 on a five-point Likert scale, with agreement levels reaching 85.71% in the highest-rated dimension. The results indicate favorable perceptions regarding the model’s consistency, interpretability, usefulness, and organizational applicability. The proposed model contributes an integrative and adaptive framework that bridges fragmented analytical practices and supports more informed, scalable, and context-aware organizational decisions.

Author 1: Juan Carlos Morales-Arevalo

Author 2: Ciro Rodríguez

Keywords: Advanced analytics; decision-making; holistic model; Design Science Research; expert judgment

Paper 55: Multi-Scale Curvelet-Based Directional Denoising for Chest X-Ray Images

Abstract: In modern healthcare, medical imaging has a significant role in understanding the structure and functioning of the human body, which helps doctors to diagnose, to plan the treatment, and to monitor the disease. Chest X-rays are widely used for the early detection and treatment of various lung infections. The effectiveness and the accuracy of the diagnosis depend largely on the quality of the medical images. Chest X-rays often suffer from Gaussian and Poisson noise, which affects the visibility of fine anatomical structures. Although methods like Gaussian filtering, NLM, and GAN have been used, they often compromise between denoising and retaining edge details. A robust denoising algorithm, Multi-Scale Curvelet Filtering with Directional Denoising (MCF-DD), is proposed to denoise medical chest X-ray images, which uses the curvelet transform coefficients. The performance of the proposed MCF-DD model was evaluated on the Chest X-Ray Images dataset from the Kaggle repository and DICOM images from the MIDRC-RICORD-1C dataset. MCF-DD achieved a PSNR of 36.57dB and SSIM of 0.9062 on Kaggle images, and 40dB PSNR with 0.9412 SSIM on DICOM images, indicating strong denoising performance across both datasets.

Author 1: Neenu Sebastian

Author 2: B. Ankayarkanni

Keywords: Image denoising; medical image; Chest X-Ray; Poisson noise; Gaussian noise; Curvelet Transform

Paper 56: Characterizing Failure Points in Rule-Based Spreadsheet Data Transformation: A Stage-Oriented Taxonomy and Empirical Failure Matrix

Abstract: Rule-based spreadsheet data transformation remains widely used for converting human-centred spreadsheet tables into structured formats for reporting, analytics, and data integration because it is transparent, auditable, and reproducible. However, rule-based approaches often fail when spreadsheets encode structural meaning through visual or layout cues, such as multi-row headers, merged cells, interleaved subtotals, cross-tabulated values, and multiple logical tables within a worksheet. This study characterizes such failures using a stage-oriented taxonomy aligned with a six-stage transformation pipeline: ingestion, region detection, structural interpretation, semantic labeling, canonicalization, and validation. A deterministic baseline pipeline was evaluated against six controlled spreadsheet cases representing common semi-structured layouts. The resulting empirical failure matrix identified 15 severity-2 failures across the controlled cases. The most critical failures were concentrated in merged cell and hierarchy handling, multi-region and boundary interpretation, role mislabeling, wrong granularity, and silent semantic errors. These findings show that many transformation failures are not isolated rule defects but arise from missing pipeline capabilities, particularly explicit hierarchy representation, block-level segmentation, granularity control, and semantic invariant checking. The study contributes a reproducible failure-analysis framework that supports stage-localized diagnosis, comparison of rule-based spreadsheet transformation systems, and targeted improvement of transformation pipelines.

Author 1: Fakhrul Adli Mohd Zaki

Author 2: Mustafa Man

Author 3: Mohamad Nor Hassan

Keywords: Rule-based transformation; spreadsheet table understanding; data wrangling; failure taxonomy; failure matrix; canonicalization

Paper 57: Arabic Sign Language Alphabet Recognition Using Transfer Learning: Evaluation, Ablation, and Deployment

Abstract: Arabic Sign Language (ArSL) is one of the most widely used sign languages among the hearing-impaired community in Arabic-speaking regions. Yet, the automated recognition of its alphabet remains a critical challenge for assistive technology development. This study presents a transfer learning classification model for Arabic Sign Language Alphabets (ArSLA) based on the InceptionV3 convolutional neural network architecture. The research contribution is a rigorously evaluated, reproducible recognition pipeline that advances diagnostic depth beyond prior studies through ablation testing, per-class F1-score analysis, and confusion pattern interpretation. The ArSL2018 dataset, comprising 54,049 images distributed across 32 Arabic alphabet classes, was used for training and evaluation. The model integrates transfer learning from ImageNet-pretrained weights, Global Average Pooling, a 256-unit dense layer with ReLU activation, dropout regularization (rate: 0.4), and a 32-class softmax output layer. Training employed the Adam optimizer with adaptive learning rate scheduling and early stopping callbacks. Model evaluation was conducted using a stratified 80:20 train-test split replicated across five independent runs with different random seeds, yielding a mean test accuracy of 97.41% ± 0.31% and a best single-run test accuracy of 97.68%, outperforming all previously reported models on the same benchmark dataset. An ablation study confirmed the independent contributions of transfer learning, data augmentation, and dropout regularization. A real-time prototype was implemented using OpenCV at 3.06 FPS on CPU hardware. These findings establish InceptionV3-based transfer learning as a strong and reproducible baseline for Arabic sign language assistive technology.

Author 1: Abdelfatah Maarouf

Author 2: Otman Maarouf

Author 3: Abdelaali Benaiss

Author 4: Rachid El Ayachi

Author 5: Mohamed Biniz

Keywords: Arabic Sign language alphabets; deep learning; transfer learning; inceptionv3; ArSL2018; gesture recognition; hearing impairment

Paper 58: Fish Disease Detection Using Modified Haar Wavelet with Adaptive Coefficient Selection

Abstract: White Spot Disease (WSD) is a major threat to aquaculture production and requires accurate image-based detection methods for early diagnosis. However, disease marker detection in fish images is challenging due to noise, illumination variations, low contrast, and complex background structures. This study proposes a Modified Haar Wavelet framework that integrates adaptive coefficient selection, Roberts operator-based edge enhancement, and hysteresis thresholding for White Spot Disease marker detection. The proposed framework aims to improve edge representation and disease marker localization while preserving disease-related structural information. Performance evaluation was conducted by comparing the proposed method against Otsu, Canny, Adaptive Thresholding, and conventional Haar Wavelet approaches using IoU, Precision, Recall, F1-score, PSNR, and SSIM. The proposed framework achieved the best overall performance with an IoU of 0.819, a precision of 0.849, a recall of 0.799, an F1-score of 0.823, a PSNR of 25.390 dB, and an SSIM of 0.889. Comparative analysis and ablation study further confirmed the effectiveness of adaptive coefficient selection, Roberts operator-based edge enhancement, and hysteresis thresholding. The results demonstrate that the proposed Modified Haar Wavelet framework provides an effective and robust solution for automated detection of White Spot Disease markers in aquaculture images.

Author 1: Tri Handayani

Author 2: Nor Hazlyna Binti Harun

Keywords: Fish disease detection; white spot disease; image analysis; modified Haar wavelet; adaptive coefficient selection; edge detection

Paper 59: Adaptive Neuro-Digital Twin with Cross-Domain Multimodal Representation Learning for Early Alzheimer's Disease Prognosis

Abstract: Alzheimer's disease (AD) refers to a progressive neurodegenerative disease involving cognitive impairment, brain atrophy, and functional neurological deficits that make early prediction and subsequent disease progression monitoring extremely difficult. Currently available AI methods predominantly focus on employing single modality analysis or static multimodal analysis approaches, which tend to solve AD prognosis as a binary classification problem. Also, much of the current literature on AD does not take into account the importance of progression-aware, patient-wise evaluation, and robust management of multimodality data heterogeneity or missingness. To tackle such issues, this study presents the proposal of an Adaptive Neuro-Digital Twin-based framework for predicting Alzheimer's disease, known as ANDT-AD. The proposed framework incorporates the heterogeneous multimodality information in clinical-cognitive data, structural Magnetic Resonance Imaging (MRI), and Electroencephalography (EEG) signals through modality-wise deep encoders. Specifically, a clinical encoder based on a transformer structure is utilized to capture non-linear cognitive interactions, while a Vision Transformer model and an attention-enhanced temporal EEG encoder model help to extract neuroanatomical information from MRI signals and electrophysiological signals in EEG, respectively. The framework is developed using Python and assessed using public domain clinical, MRIs, and OpenNeuro ds004504 EEG data sets for five-fold cross-validation and patient-level evaluations. Experimentation yielded 98.0% diagnostic accuracy with an AUC of 0.97, which exceeds the performance of current multimodal baselines. Moreover, the framework attained an MAE of 1.12, an RMSE of 1.46, and a progression risk C index of 0.89, proving its strength in predicting cognitive decline and personalizing disease progression models under heterogeneity and missingness of multimodal scenarios.

Author 1: V S Krushnasamy

Author 2: Annapurna Mishra

Author 3: Pratik Gite

Author 4: Ganesh Kumar Anbazhagan

Author 5: Adapa Gopi

Author 6: M.Misba

Author 7: A. Arul Anitha

Author 8: Osama R.Shahin

Keywords: Alzheimer’s disease; neuro-digital twin; multimodal learning; disease prognosis; cross-domain representation learning

Paper 60: Semantic Style Transfer for Paintings Using Convolutional Neural Networks (CNNs)

Abstract: In recent years, the importance of photographic portrait styles has garnered significant attention, prompting numerous researchers to explore innovative methods for modifying and enhancing these styles. Neural style transfer has advanced rapidly for photographic portraits, yet transferring painterly styles to human facial and body images remains difficult because global stylization frequently distorts facial geometry and erases identity. This study presents a semantic, region-wise painting style transfer framework based on Convolutional Neural Networks (CNNs) that preserves facial identity and semantic structure during stylization. The method parses both the source photograph and an example painting into corresponding semantic regions, comprising ten facial components together with hair, chest, arms, legs, and background, and transfers the style region by region so that each part is stylized from its semantic counterpart. A feature reconstruction stage based on Gram matrix style representations minimizes content and style loss within each region, while a part-based generation and fusion stage augmented with Laplacian pyramid decomposition improves local to global consistency and identity preservation. We evaluate the approach with perceptual and identity metrics, reporting Frechet Inception Distance (FID), Structural Similarity (SSIM), and identity cosine similarity (CSIM), and additionally report a downstream classification check as an auxiliary indicator of content preservation. The full model attains an FID of 14.72, an SSIM of 0.82, and a CSIM of 0.86, outperforming GAN and part generation network baselines in identity preservation and realism. We discuss the strengths, limitations, and practical implications of the framework and outline directions toward full-body, high-resolution, and video stylization.

Author 1: Hafiz Muhammad Jamsheed Nazir

Author 2: Zheng Jiangbin

Author 3: Omar Alsaleh

Keywords: Painting style transfer; semantic style transfer; identity preservation; facial region parsing; Convolutional Neural Networks (CNNs)

Paper 61: Multi-Dimensional Fractal Vulnerability Study Of Alzheimer's Brain Networks

Abstract: Alzheimer’s disease (AD) is associated with widespread disruption of functional brain networks. While graph-theoretic studies have characterized altered connectivity patterns in Alzheimer’s disease, most of the analyses rely on single-scale representations and fixed threshold choices, therefore obscuring critical structural transitions. In this study, the multi-scale fractal vulnerability of EEG-derived functional networks in Alzheimer’s disease is investigated using a density-controlled graph framework. Resting-state EEG recordings from AD subjects are converted into functional connectivity networks, systematically sparsified across a range of network densities, and analyzed using the box-counting fractal dimension (FD), restricted to the largest connected component (LCC). Fractal structure in AD networks degrades non-monotonically with decreasing density, and a transitional density regime (~15-25%) was identified within the analyzed Alzheimer’s disease cohort, corresponding to a regime of heightened structural vulnerability. These findings accentuate the importance of multi-scale analysis for understanding brain network organization.

Author 1: Tin Tin Ting

Author 2: Neha Hema Raj

Author 3: Geetha N K

Author 4: Thangaraj C

Author 5: Olusegun D. Samuel

Keywords: Fractal dimension; graph theory; Alzheimer’s disease; EEG; transitional density regime; functional connectivity; human health

Paper 62: Progression-Aware Temporal Graph Transformer for Reliable Chronic Kidney Disease Trajectory Prediction

Abstract: Chronic Kidney Disease (CKD) is a progressive and irreversible condition that requires early prediction of renal deterioration for effective clinical intervention. Existing studies based on static machine learning and conventional deep learning models fail to capture temporal dependencies, evolving biomarker interactions, and longitudinal disease progression patterns, leading to limited predictive reliability in real-world clinical settings. To address these limitations, this study proposes a Progression-Aware Temporal Graph Transformer for reliable CKD trajectory prediction using the Chronic Renal Insufficiency Cohort (CRIC) dataset comprising 5,625 patients with longitudinal clinical follow-ups and repeated biomarker measurements including eGFR, creatinine, albumin, blood pressure, glucose, and HbA1c. The proposed framework integrates a temporal transformer encoder for sequential patient representation, a dynamic biomarker graph to model evolving renal interactions, and a multi-scale temporal attention mechanism to capture both short- and long-term deterioration patterns. The model is implemented using Python with PyTorch and executed on GPU-based computational infrastructure for efficient training and inference. Experimental results demonstrate that the proposed model achieves 95.0% accuracy, 94.2% precision, 95.1% recall, and 94.6% F1-score, improving performance by approximately 2–4% over state-of-the-art baselines such as Random Forest, KNN, GNN, and 1D CNN models. Additionally, survival analysis yields a Concordance Index of 0.952, confirming strong risk-ranking capability for dialysis onset prediction. The framework also maintains robust performance under noisy and missing data conditions, demonstrating strong generalization ability. In conclusion, the proposed model provides an interpretable, scalable, and clinically reliable solution for CKD progression forecasting, enabling early intervention, personalized treatment planning, and improved renal outcome prediction in clinical decision-support systems.

Author 1: Roshan D. Suvaris

Author 2: Padmavathy E

Author 3: Dilfuza Akabirkhodjaeva

Author 4: T. K. Rama Krishna Rao

Author 5: R. Sindhu

Author 6: Farrukh Sobia

Author 7: Elangovan Muniyandy

Author 8: Aseel Smerat

Keywords: Chronic Kidney Disease; temporal graph learning; disease progression prediction; longitudinal healthcare modeling; trajectory prediction

Paper 63: Toward Adaptive Educational Intervention: Meta-Adaptive Cross-Modal Gating for Few-Shot Personalized Intervention

Abstract: The rapid evolution of AI-enhanced learning environments has created an urgent need for intelligent educational systems capable of delivering early and personalized interventions under severe data sparsity conditions. This study proposes a Meta-Adaptive Cross-Modal Gating (MACMG) mechanism integrated within a Multimodal Transformer-based Educational Digital Twin framework for early detection of at-risk students. The proposed approach addresses the cold-start problem by introducing a student-specific, meta-learned gating policy that dynamically fuses textual, behavioral, and physiological educational signals. At the core of MACMG lies a bilevel optimization strategy inspired by Model-Agnostic Meta-Learning (MAML), enabling rapid adaptation to unseen learners using only a few initial interaction episodes. The framework generates time-varying modality weights that emphasize informative signals while suppressing noisy channels. To preserve representational capacity and computational efficiency, the adaptive gating mechanism is integrated only into the top two layers of a six-layer multimodal Transformer using a first-order MAML approximation. The personalized representations are propagated toward a risk prediction module and a digital twin state manager responsible for continuously updating learner knowledge states. Experimental results demonstrate that MACMG achieves AUROC scores of 0.821 on StudentLife and 0.834 on DAiSEE, outperforming static multimodal Transformers, conventional fine-tuning approaches, and full-model meta-learning baselines. Furthermore, the proposed framework reduces learner-specific adaptation time to only 0.47 seconds while maintaining robust performance under sparse-data conditions, highlighting its suitability for real-time personalized educational intervention.

Author 1: Houda Kaa

Author 2: Hanane Allioui

Author 3: Ilham Oumaira

Keywords: Multimodal transformer; meta-learning; at-risk student detection; personalized learning; educational data mining; adaptive intervention

Paper 64: An Adaptive Smart Business Intelligence Model Based on Enhanced HGO Discovery for Real-Time in Inpatient Care

Abstract: The complexity and dynamics of inpatient care require advanced decision support systems that are fast, adaptive, and capable of real-time execution. This study proposes a novel Smart Business Intelligence (SBI) model developed through an enhanced Hierarchy Governance Outlook (HGO) Discovery approach to achieve high-precision inpatient service prioritization. Unlike conventional frameworks, the proposed model combines an adaptive weighting mechanism, a sustainability optimization process, and real-time data integration from heterogeneous clinical and operational sources. The enhanced HGO Discovery model allows for dynamic adjustment of decision parameters in response to evolving hospital conditions, thereby maximizing patient care. The results demonstrate that the proposed model consistently outperforms baseline approaches in predictive accuracy, stability, and computational efficiency. These findings highlight the potential of the proposed framework to support data-driven decision-making and enhance the quality and efficiency of inpatient care delivery in modern healthcare environments. Experimental results show that the proposed model provides more accurate patient prioritization, faster identification of critical conditions, and improved operational efficiency compared to conventional approaches. Therefore, the proposed system offers an effective intelligent healthcare solution to support real-time inpatient care and data-driven medical decision-making in modern hospitals.

Author 1: Hengki

Author 2: Rahmat Gernowo

Author 3: Oky Dwi Nurhayati

Keywords: Smart business intelligence; decision support systems; HGO discovery; inpatient care; hospital

Paper 65: Multi-Relation Knowledge Graph-Guided Transformer with Objective-Aware Route Selection for Vehicle Routing Problems with Time Windows

Abstract: Vehicle routing problems with time windows require route construction methods that can reason jointly about spatial distance, vehicle capacity, service time, and customer time windows. Recent neural combinatorial optimization studies show that Transformer and graph-based models can learn routing policies, but many models still rely mainly on node features and do not expose which routing relations are responsible for feasibility and route quality. This study proposes a multi-relation knowledge graph guided Transformer framework for the vehicle routing problem with time windows. For each instance, a pairwise knowledge graph is built from spatial, temporal, capacity, slack, and depot identity relations. These relation vectors are projected into head-specific attention biases inside a pointer-style Transformer, while a feasibility mask enforces capacity and time window constraints during decoding. The framework is trained by imitation from OR-Tools teacher routes and evaluated on the 56 Solomon benchmark instances and on synthetic Solomon-style extensions. The original Transformer baseline decoded feasible routes for 54 of 56 Solomon instances, whereas the full knowledge graph guided model decoded feasible routes for all 56 instances. Additional controlled AM style, POMO style multistart, and edge-aware Transformer baselines improve route quality gaps on feasible instances but still leave one Solomon instance infeasible, highlighting the proposed framework's feasibility-oriented role. On three 120 instance synthetic runs, objective-aware route selection over relation-specialized models reached a mean feasible rate of 0.997 and reduced the mean distance gap by 125.418 Solomon coordinate distance units compared with the relation-free Transformer baseline. The contribution is therefore a diagnostic and extensible framework that integrates explicit routing knowledge into neural attention, evaluates relation ablations, and combines specialized route candidates according to the target operational objective.

Author 1: Somkiat Kosolsombat

Author 2: Chiabwoot Ratanavilisagul

Keywords: Vehicle routing problem with time windows; knowledge graph; Transformer; neural combinatorial optimization; relation aware attention; route selection

Paper 66: Deployment-Aware 30-Day Readmission Prediction in Resource-Limited Hospitals: Calibration, Threshold Policy, and Decision Utility

Abstract: Thirty-day hospital readmission is a well-established quality metric, and many clinical prediction models have been developed for this task; however, high discrimination does not by itself mean that a model is safe to use in discharge workflows. This study developed and applied an integrated deployment-oriented evaluation workflow in which calibration, threshold governance, and decision utility were treated as primary evaluation requirements rather than secondary diagnostics. Retrospective inpatient data collected between 2022 and 2024 from two resource-limited government hospitals were used (N = 30,000; readmission prevalence = 15.0%). The analysis was based on patient-level internal validation using non-overlapping training, validation, and held-out test partitions. A multilayer perceptron neural network and a random forest were evaluated using patient-level grouping (70% training, 15% validation, 15% test). Both models showed strong discrimination on the held-out test set (ROC-AUC = 0.868 for the neural network and 0.880 for the random forest), with nearly identical minority class detection (PR-AUC = 0.461 vs 0.460). However, calibration analyses separated the models despite identical Brier scores (0.095). The neural network showed lower expected calibration error (ECE = 0.013 vs 0.035) and near ideal probability scaling (calibration slope = 0.971, 95% CI: [0.94, 1.00]) compared with the random forest (slope = 1.202). Threshold analysis also showed that a default threshold could be unsafe, since recall was 0.244 at 0.50 but increased to 0.867 at 0.12, while false negatives dropped from 510 to 90. Decision Curve Analysis further supported the neural network, including a mean net benefit of 0.097 at a threshold of 0.12. Practically, threshold, model version, and monthly calibration summaries should be logged in an audit trail.

Author 1: Samer Asad Malalha

Author 2: Ma Burhanuddin

Author 3: Hatem T M Duhair

Author 4: Jamil Abedalrahim Jamil Alsayaydeh

Author 5: Mazen Farid

Keywords: Clinical AI deployment; hospital readmission prediction; expected calibration error; electronic health records; operating region; probability reliability; model governance

Paper 67: Expert Evaluation of the Proposed ERP Model for Legal Compliance and Adaptation

Abstract: This study presents an expert-based evaluation of a proposed context-aware ERP model designed to support adaptive legal compliance in multilingual regulatory environments, with a focus on North Macedonia. The model integrates modular ERP architecture, AI-driven legal reasoning, multilingual natural language processing, adaptive learning, and human oversight and is intended to support continuous alignment with changing legislation. A structured questionnaire was administered to 51 experts from fields, including ERP systems, software engineering, artificial intelligence, legal compliance, and public administration, to assess the model’s conceptual clarity, technical feasibility, multilingual support, scalability, and compliance functionality. The findings provide expert-based conceptual support for the proposed framework, particularly regarding its relevance, architectural coherence, and potential applicability, but they should not be interpreted as evidence of operational effectiveness in a deployed ERP environment. Participants particularly emphasized the importance of multilingual support and human validation mechanisms for maintaining accountability and trust in AI-assisted compliance processes. The evaluation also identified implementation challenges related to integration complexity, AI interpretability, and organizational readiness, highlighting areas for future prototype development and empirical validation.

Author 1: Shkëlqim Miftari

Author 2: Azir Aliu

Author 3: Artan Luma

Keywords: ERP systems; legal compliance; artificial intelligence; multilingual NLP; adaptive learning; expert evaluation; compliance automation

Paper 68: AI-Driven Service Innovation, Customer Satisfaction, and Guest Loyalty: Evidence from Shenyang Airport Hotel, China

Abstract: This study investigated how guests' technological perceptions of Artificial Intelligence (AI) applications perceived usefulness (PU), perceived ease of use (PEOU), enjoyment (ENJ), and privacy concerns (PC) influence customer satisfaction (CS) and customer loyalty (CL) at the Shenyang Airport Hotel, a three-star airport property in a secondary Chinese city. Grounded in an integrated framework combining the Technology Acceptance Model (TAM) with Customer Experience Theory. A quantitative survey of 452 guests who had experienced AI-enabled services during their stay was conducted. Data were analysed using reliability assessment, exploratory factor analysis (to verify the factor structure in a novel context), Pearson correlation, multiple regression, and bootstrap mediation analysis. The results indicated that perceived usefulness (β = 0.314, p < 0.001), enjoyment (β = 0.141, p < 0.01), and perceived ease of use (β = 0.127, p < 0.01) each positively influenced customer satisfaction, whereas privacy concerns exerted a significant negative effect (β = −0.205, p < 0.001). Customer satisfaction, in turn, significantly predicted customer loyalty (β = 0.409, p < 0.001), and partially mediated all four perception-to-loyalty pathways. One-way ANOVA further revealed significant differences in customer loyalty across age groups, educational levels, and income brackets. Theoretically, this study extended TAM by simultaneously incorporating affective and security-related dimensions into AI service acceptance within an underexamined hospitality context. Practically, the findings offer airport hotel managers in secondary cities evidence-based guidance on prioritising utility-enhancing AI features, simplifying service interfaces, enriching hedonic engagement, and communicating data governance policies transparently to mitigate privacy-driven dissatisfaction. The study was limited by its single-site, cross-sectional design; future research should adopt longitudinal or multi-site approaches to strengthen generalisability.

Author 1: Dayong Zu

Author 2: Kawalin Angkananon

Author 3: Yoksamon Jeaheng

Keywords: Artificial intelligence; airport hotels; technology acceptance model; customer satisfaction; customer loyalty; secondary city; China

Paper 69: Ontology-Based Business Process Modeling: A Review

Abstract: Business Process Modeling (BPM) has been receiving attention in recent years. Organizations operating in distributed, data, and knowledge-intensive environments need precise machine-interpretable process descriptions. Traditional business process modeling notations such as BPMN, UML Activity Diagrams, and EPCs are highly effective for visualizing workflows and supporting communication among stakeholders but do not address problems such as semantic ambiguity, inconsistent terminology, limited reuse, and poor interoperability across organizational boundaries. Since ontology-based models can enable seamless process integration, coordination, and collaboration among autonomous systems. Therefore, ontology-based descriptions are well-suited for complex enterprise systems and supply chains. Ontology-based business process modeling techniques provide a robust and theoretically grounded solution by introducing explicit semantics, formal reasoning capabilities, and shared conceptual frameworks into business process models. This semantic enrichment significantly enhances the expressive power of business process descriptions while preserving compatibility with existing modeling standards. However, ontology-based business process modeling also has its own challenges. The development and maintenance of high-quality ontologies require significant effort and domain expertise. Moreover, reasoning over large-scale ontologies may introduce computational overhead, particularly in real-time environments. Tool support is another practical concern, as seamless integration between business process modeling tools and ontology management platforms is still limited in many industrial settings. The primary objective of this study is to conduct a systematic literature review of ontology-based business process modeling approaches to provide research recommendations based on their strengths and limitations. The results indicate that there are several research gaps that should be addressed to ensure smooth process integration across organizational boundaries. Additionally, empirical validation of ontology-based BPM frameworks in real-time environments is limited. The proposed framework not capable enough to update ontologies with the evolution of the business and related processes. In this regard, an Input Process Output (IPO) BPM Framework to integrate ontologies into BPM as a three-stage transformation mechanism is proposed.

Author 1: Low Kok Thai

Author 2: Furkh Zeshan

Author 3: Nazri Kama

Author 4: Riza Sulaiman

Author 5: Mohammad Nazir Ahmad

Author 6: Uwais Qidwai

Keywords: Business Process Modeling (BPM); Ontology-based Business Process Modeling (OBPM); domain modeling; knowledge management; knowledge reasoning; knowledge engineering; knowledge modeling

Paper 70: A Robust Security Framework for Cloud Data Storage Using Lightweight Blockchain Technology

Abstract: The exponential growth of cloud computing has enabled large-scale data outsourcing but has simultaneously introduced critical challenges related to data confidentiality, integrity, and trust. Traditional cryptographic and blockchain-based cloud security solutions often suffer from high computational overhead, latency, and scalability limitations, which hinder their practical adoption. To address these issues, this study proposes a robust and lightweight blockchain-based security framework for secure cloud data storage. The framework integrates hybrid AES–ECC encryption, smart contract–driven access control, and a lightweight consensus mechanism combining Delegated Proof of Stake (DPoS) and Practical Byzantine Fault Tolerance (PBFT) to achieve efficient and tamper-resistant data management. The proposed system employs an on-chain/off-chain hybrid architecture that stores only essential metadata and cryptographic proofs on the blockchain while maintaining the actual data in distributed cloud storage. This design minimizes computational burden and blockchain bloat while ensuring end-to-end transparency and verifiability. A Merkle tree–based Proof of Storage (PoS) mechanism enables rapid integrity verification without requiring full data retrieval. Comprehensive experiments were conducted using a simulated multi-node cloud environment to evaluate encryption efficiency, transaction latency, throughput, storage overhead, and energy consumption. Results show that the proposed framework outperforms existing blockchain-based models, achieving a 37.7% reduction in encryption/decryption time, a 51.3% decrease in transaction latency, and a 54.5% improvement in energy efficiency. Additionally, the system attained a 99.3% security success rate under various attack scenarios, demonstrating its resilience against unauthorized access, replay, and tampering attempts. These findings confirm that the proposed approach provides a practical balance between security assurance and performance optimization.

Author 1: Renuka GOLLA BALA

Author 2: S. Gnanavel

Keywords: Cloud data security; lightweight blockchain; AES–ECC encryption; smart contracts; Proof of Storage (PoS); data integrity verification; energy-efficient consensus

Paper 71: ERP Implementation Success Factors Across Project Phases: An Action Research and Fuzzy AHP Study in Moroccan SMEs

Abstract: Today, adopting an ERP has become a critical decision for Small and Medium-sized Enterprises (SMEs) seeking to modernize their information systems and implement integrated management. In this context, it is crucial to identify the Critical Success Factors (CSFs) to ensure the successful ERP Implementation (ERPI). However, previous studies have generally analyzed these factors in a broad sense, without considering their progression through the various phases of the ERPI process. The objective of this study is to identify the CSFs associated with each phase of the ERPI process, as well as to determine their relative importance in each of these phases within Moroccan SMEs. To achieve this objective, a mixed-methods approach was adopted. The action research method was used to identify the CSFs associated with each phase of the ERPI process within Moroccan SMEs. Additionally, we used the Fuzzy AHP method to determine their relative importance in each of these phases. The results indicate that the importance of CSFs varies across the different phases of ERPI. During the analysis phase, strategic factors are paramount, while those related to project management and technology gain in importance during the design and implementation phases. Finally, organizational factors play a crucial role during the final preparation phase, as well as during the implementation and support phase. This research complements the existing literature, which analyzes key success factors in a general way, by offering a phase-by-phase analysis of the ERPI process. It provides ERP consultants and project managers with a clear overview of the factors to prioritize at each phase of the project, and helps reduce the risk of ERP failure by identifying the key success factors that ensure the successful implementation of ERP within Moroccan SMEs.

Author 1: Yassine Zouhair

Author 2: Younous Elmrini

Keywords: ERP; ERP Implementation; CSFs; Moroccan SMEs; Integrated IS

Paper 72: Hybrid Data Fusion and Deep Learning for Dynamic Risk Index Modeling in Secure Learning Management Systems

Abstract: The explosive growth of online education platforms has led to increased exposure to cybersecurity threats, which makes secure Learning Management Systems (LMS) a critical requirement. However, the current methods often can't capture user behavior risk and network-level attack patterns at the same time, which causes the threat to be incomplete. This study presents a dynamic cyber risk prediction model by fusing log information of LMS behavior with network intrusion information in the CICIDS2017 dataset. The goal is to create an AI-based model that is able to perform real-time risk assessment using a Dynamic Risk Index (DRI). The methodology includes the combination of feature engineering, hybrid data fusion, machine learning, deep learning (LSTM, DNN), and anomaly detection methods. Experimental results demonstrate that the proposed model achieves an accuracy of 97.6%, an F1-score of 96.9%, and an AUC of 98.5%, outperforming state-of-the-art methods. The robustness and significance of the framework are confirmed by ablation and statistical analyses. The overall study concludes that combining behavioral and network intelligence with dynamic risk scoring improves cyber threat detection and proactive security management in e-learning environments.

Author 1: Vani T

Author 2: S. Sathya

Keywords: Cybersecurity; Dynamic Risk Index (DRI); e-learning security; behavioral analytics; intrusion detection; hybrid deep learning; LMS logs; CICIDS2017 dataset; anomaly detection; risk prediction

Paper 73: Adaptive Honey Badger Optimization with Bernoulli Chaotic Mapping and Decreasing Neighborhood for Efficient Cloud Task Scheduling

Abstract: One of the most important challenges in optimizing task scheduling in cloud computing is the dynamic nature of resources, the heterogeneity of tasks, and conflicting optimization criteria, such as minimizing task completion period, optimizing resource utilization, and shortening task migration time. Meta-heuristics such as the Honey Badger Algorithm (MBA) often converge on suboptimal solutions and may not strike the perfect balance between exploitation and exploration in task scheduling optimization. To address the problems associated with traditional HBA and similar algorithms, this research introduces a novel optimization technique called the Multi-strategy Honey Badger Algorithm (MHBA). The proposed MHBA integrates three optimization strategies: horizontal crossing coupled with adaptation, an optimum decreasing neighborhood, and a Bernoulli shift scheme. The MHBA is simulated using CloudSim and compared with other advanced techniques. The experimental findings confirm MHBA's efficacy in reducing makespan by up to 25.6%, increasing resource utilization by up to 16.7%, and decreasing migration time by up to 27.5% when applied to the HPC2N dataset. The same was evident in the NASA dataset, where MHBA achieved reductions in makespan of up to 25.9%, improvements in resource utilization of up to 21.6%, and reductions in migration time of up to 24.1%.

Author 1: Yanfang XING

Keywords: Cloud computing; multi-objective task scheduling; honey badger; Bernoulli chaotic mapping; exploration–exploitation balance

Paper 74: Structure-Aware Latent Diffusion for High-Quality Line Art Colorization

Abstract: To address the limitations of existing line art colorization methods in structural preservation, color mapping accuracy, and semantic consistency, this study proposes a structure-aware multi-instance constrained line art colorization method based on latent diffusion. Built upon the latent diffusion framework, the proposed method introduces a structure-aware constraint mechanism to enhance the preservation of line contours and edge details during generation. Meanwhile, instance-level semantic modeling and feature fusion strategies are incorporated to achieve coherent local color representation and optimize global semantic consistency. In addition, a unified optimization objective is constructed by jointly integrating structural constraints, color consistency constraints, and regularization terms, thereby improving the visual quality and naturalness of the generated results through collaborative multi-constraint learning. Experimental results on public datasets demonstrate that the proposed method outperforms comparative approaches in terms of FID, PSNR, SSIM, and LPIPS, producing high-quality colorization results with clear structures, natural colors, and strong semantic consistency, which verifies its effectiveness and superiority.

Author 1: Shuhua Xu

Author 2: Qiang Ai

Author 3: An Zhao

Author 4: Guan Yang

Author 5: Bo Chen

Keywords: Line art colorization; latent diffusion; structure-aware modeling; instance-level semantic modeling; feature fusion; image generation

Paper 75: C-TriLoRA: Cross-Corpus Kazakh SER via Tri-Factor LoRA and CORAL

Abstract: Speech Emotion Recognition (SER) in low-resource languages deals with the scarcity of labeled corpora and the instability of learned representations when transferred across diverse recording conditions and speaker demographics. This study introduces Conditional Tri-Factor Low-Rank Adaptation (C-TRILORA), a multi-task architecture that jointly performs automatic speech recognition (ASR) and SER on Kazakh speech while generalizing reliably across corpora. The proposed model extends a pre-trained Whisper encoder–decoder back-bone through three primary innovations: a Tri-LoRA routing module that disentangles lexical, emotional, and speaker latent factors; CORAL domain alignment that matches second-order statistics between source and target domains without target labels; and a gradient reversal layer (GRL) that suppresses speaker-identity information. Experimental evaluations on the KazEmoTTS and ENU KEMO datasets demonstrate that C-TRILORA achieves a competitive in-domain Macro-F1 of 86.09%and significantly outperforms standard baselines in cross-corpus conditions (41.44% Macro-F1 versus 37.34% for the Shared-Head baseline). McNemar and Wilcoxon signed-rank tests confirm that explicit factor disentanglement is essential for cross-corpus robustness. These results show that separating speech components effectively mitigates negative transfer, making C-TRILORA a practical approach for low-resource SER deployment.

Author 1: Bakdaulet Kynabay

Author 2: Aimoldir Aldabergen

Author 3: Shirali Kadyrov

Keywords: Speech emotion recognition; low-resource languages; parameter-efficient fine-tuning; domain adaptation; multi-task learning; disentangled representations

Paper 76: Query Recovery Attack Based on Multi-Source Leakage and Semantic Embedding in Searchable Symmetric Encryption

Abstract: With the rapid development of cloud computing and big data technologies, searchable encryption has become a research hotspot. To improve search efficiency, searchable encryption algorithms may leak redundant information, allowing adversaries to launch query recovery attacks by exploiting pattern leakage in searchable symmetric encryption and infer the underlying keywords of user queries. Existing query recovery attack methods only perform query recovery under multi-source leakage patterns, without considering the semantic level; they simulate user queries with plaintext rather than ciphertext and rarely take frequency as auxiliary leakage for attacks. Meanwhile, the weights between volume and frequency mostly rely on manual configuration, without dynamic allocation of the weights for volume and frequency. Therefore, this study proposes a novel attack method named refine atk. First, a multi-layer perceptron is used to learn the weights of volume and frequency to accurately identify and recover distinctive queries. Next, co-occurrence information is employed to correct the queries recovered in the previous step. Finally, a cost matrix is constructed using the weighted co-occurrence matrix and the semantic embedding matrix obtained by the pre-trained language model MiniLM-L6, and the remaining queries are recovered in one pass via greedy graph matching. The proposed attack achieves an attack accuracy of 95% on the Enron and Lucene datasets. The attack performance remains robust even after removing part of the similar data. The attack execution efficiency of the proposed method is significantly superior to that of the traditional schemes, yielding better performance when attacking concrete searchable symmetric encryption schemes or evaluating their security. This work provides a reference for the security evaluation and defense mechanism design of searchable symmetric encryption.

Author 1: Xiaogang Yuan

Author 2: Xinle Yang

Author 3: Dezhi An

Keywords: Searchable symmetric encryption; query recovery attack; pattern leakage; multi-layer perceptron (MLP)

Paper 77: Multiplicative Gate State Space Models with Skip-Net for High Accuracy COVID-19 Time-Series Prediction

Abstract: The rapid propagation of the COVID-19 pandemic has placed unprecedented strain on global healthcare systems, creating an urgent need for accurate forecasting to optimize resource allocation and policy implementation. However, the highly non-linear and chaotic behavior of infection rates poses significant challenges for traditional statistical and standard deep learning models. This study proposes the Multiplicative Gating State Space Model with skip connection (MG-SSM-s), a novel architecture designed to capture complex temporal dependencies in epidemiological time series. Drawing inspiration from partial autocorrelation, the model extends modern State Space Models (SSMs) by incorporating a learnable multiplicative side channel to dynamically modulating input processing. We evaluated the efficacy of MG-SSM-s using the Google COVID-19 Open Data repository, analyzing daily confirmed cases across 40 countries, including major epicenters such as the USA, India, and Brazil. Using a 30-day look-back window, the proposed model was benchmarked against four baseline architectures: LSTM, Bi-LSTM, GRU, and standard SSM. Performance verification based on Mean Squared Error (MSE) demonstrates that MG-SSM-s outperforms all deep learning baselines and achieves competitive accuracy with a tuned ARIMA model, demonstrating comparable statistical performance to the latter. These results highlight the framework’s robustness and potential as a versatile tool for time-series forecasting.

Author 1: Krung Sinapiromsaran

Author 2: Supakit Sroynam

Keywords: State Space Model; COVID-19; time-series fore-cast; univariate time-series

Paper 78: Topology-Aware Fingerprint Representation Using Graph Convolutional Network for Robust Minutiae Refinements

Abstract: Most of the fingerprint recognition systems represent minutiae as independent points. This reduces robustness under noise, distortion, and partial impressions. The lack of explicit structural modeling contributes to inconsistent feature reliability, specifically in defocused acquisition conditions. To address these limitations, a Topology-Aware Graph Convolutional Network (Topo-GCN) is proposed. Topo-GCN is a geometric fingerprint representation framework that treats all minutiae points as interrelated nodes and polishes them using a Graph Convolutional Network (GCN). Each node is represented as a 14-dimensional descriptor with the Multi-scale Spatial Feature Tensor (MSFT) technique. A novel Relational Connection Factor (RCF) applies adaptive topology-aware learning to construct edges in place of traditional distance-based graph construction. The Hybrid scoring technique combines learned node representations with graph-theoretic measures to separate genuine minutiae from false ridge information. Furthermore, a lightweight pseudolabeling scheme efficiently trains the model without depending on the large-scale annotated datasets. The proposed Topo-GCN framework achieves an Area under Curve (AUC) of 0.9614, 0.9769, and 0.9831 with consistent verification performance (EER) of 1.85%, 1.39%, and 1.64% and stable relational encoding with mean edge weights of 0.712, 0.739, and 0.758 across FVC2000, FVC2002, and FVC2004 datasets. The results indicate that integrating relational topology into minutiae modeling considerably enhances robustness, making Topo-GCN a viable approach for secure and efficient fingerprint authentication systems.

Author 1: Yoogesh A

Author 2: Rama Prasath A

Keywords: Biometric authentication; fingerprint recognition; Graph Convolutional Network; minutiae refinement; topology-aware representation

Paper 79: Architecting a Low-Latency RAG System for Fast-Moving Consumer Goods (FMCG) Customer Support: A Case Study in Industrial Software Deployment

Abstract: Deploying large language models (LLMs) in industrial customer support environments require balancing response accuracy with system latency. This study presents the software architecture and implementation of a Retrieval-Augmented Generation (RAG) system designed for the Fast-Moving Consumer Goods (FMCG) sector. Addressing the limitations of generic LLMs in domain-specific knowledge tasks, we engineered a retrieval-augmented inference pipeline that integrates unstructured data ingestion, Pinecone for vector indexing, and Groq-based inference for low-latency response generation. The proposed system aims to improve response grounding by incorporating organizational product information into the generation process while maintaining responsive interaction times suitable for customer support applications. This study details the software architecture, system integration approach, and experimental evaluation of the proposed deployment-oriented RAG framework in an industrial FMCG case-study setting.

Author 1: Meredita Susanty

Author 2: Alghifari Rasyid Zola

Keywords: Software architecture; RAG; industrial AI; Groq LPU; vector database; customer support systems

Paper 80: Adaptive Hybrid Intrusion Detection for Realistic Zero-Day Attacks in Cloud and Edge Environments

Abstract: New and unknown attack patterns are creating more cybersecurity issues for cloud environments. Intrusion detection systems (IDS) are usually capable of high-performance in closed-world scenarios and are less effective in realistic zero-day scenarios. This work presents an adaptive hybrid intrusion detection mechanism for cloud environments based on the combination of Random Forest, XGBoost and AutoEncoder models, under a single decision-making framework. A fully reproducible pipeline was built directly from the raw data of CICIDS2017, using consistent preprocessing and feature engineering. In contrast with traditional IDS researches that are evaluated with standard train-test evaluation process, realistic zero-day assessment was carried out by conducting attack-family holdout experiments, including WebAttack, DDoS, and Infiltration scenarios. The results showed that both attack families and adaptive decision strategies were different, with WebAttack using XGBoost-dominant behavior and Infiltration using anomaly-focused behavior for an 85.30% and 88.89% recall, respectively. The proposed framework was also implemented on edge hardware (Raspberry Pi) and has been shown to have low-inference latency and only moderate resource usage. The results reveal that adaptive hybrid decision behavior leads to a more robust decision behavior under realistic zero-day conditions, while having a feasible deployment in cloud-edge environments.

Author 1: Nithin U

Author 2: Ganeshayya Shidaganti

Author 3: Sangeetha V

Author 4: Vishwachetan D

Keywords: Intrusion detection system (IDS); zero-day detection; adaptive hybrid IDS; attack-family holdout; cloud security; edge deployment; autoencoder; CICIDS2017

Paper 81: Towards Reliable Recognition of Concurrent Abnormal Patterns in Control Charts Using Multi-Label Deep Learning

Abstract: Control charts do more than raise an alarm: their shapes can give an early indication of what has changed in a process. This study considers the case in which one chart window contains more than one abnormal behavior. The observed sequence is then a mixture rather than a pure pattern. We formulate this problem directly as multi-label classification. A one-dimensional CNN receives a raw-scale window of 32 observations and predicts the active elementary labels. The controlled protocol contains twelve scenarios: normal behavior, six single abnormal patterns, and five selected concurrent patterns. Raw-scale input is retained because shift patterns depend partly on level information that may be weakened by window-wise normalization. The retained training setup gives additional exposure to difficult shift and trend cases, while validation and testing remain balanced. Across five repeated trainings, the model achieved 96.11% exact match accuracy, 96.41% precision, 96.46% recall, 96.44% F1-score, and 1.04% Hamming loss. The 95% confidence interval for exact match was 96.05–96.17%. Additional analyses show stable performance around a decision threshold of 0.5, strong cyclic and systematic recognition, and lower performance for short shift cases. The results support direct multi-label CNN recognition for the selected protocol. Broader shift-containing mixtures, more complex combinations, varying noise conditions, and real industrial validation remain outside the scope of the present controlled study.

Author 1: Mohammed Modar

Author 2: Abdelilah Ganmati

Author 3: Omar El Farissi

Keywords: Control chart pattern recognition; concurrent patterns; statistical process control; multi-label classification; convolutional neural network; process monitoring

Paper 82: Deep Learning for the Classification of Kidney Diseases in Medical Images Using ResNet-50 and Grad-CAM

Abstract: Renal pathology represents a diverse set of diseases that present significant clinical relevance. Included among the various types of renal pathologies are renal stones, cysts, and renal malignancies, all of which require diagnosis and therapy to prevent progression of the disease process. The current research study was performed to create and validate a classification model based on deep learning using a convolutional neural networks (CNN) architecture, namely a 50-layer Residual Network (ResNet-50) using Gradient-weighted Class Activation Mapping (Grad-CAM), to provide improved automatic detection of renal pathology from medical images and improve the interpretability of those medical images. During the study, the Explainable Deep Learning Pipeline (X-DLP) paradigm was followed, which provides a structured methodology to perform research with the use of deep learning in medical imaging. The X-DLP structures the research process into a series of phases, including Data acquisition and curation, Preprocessing and Augmentation, Model Creation via Transfer Learning, and lastly, Interpretability and Visualization.The results obtained show that the proposed model performs consistently well across different evaluation metrics. The Precision–Recall curve, with a PR-AUC close to 0.89, suggests that the model is effective at identifying positive cases even when the data are imbalanced. In addition, the F1-score reaches a peak of around 0.835 at a threshold near 0.45, indicating a good trade-off between precision and recall. From another perspective, the evaluation using Youden’s criterion reveals sensitivity and specificity values close to 0.80, which supports the model’s ability to distinguish between classes with reasonable accuracy. Moreover, the lift and cumulative gain analysis further highlight its practical usefulness, with a lift of 3.5 in the top 10% and a cumulative gain of 75% when considering 30% of the population. These results indicate that the model can effectively prioritize the most relevant positive cases. Overall, these findings suggest that the model can serve as a valuable support tool in medical diagnosis. By enabling automated classification of renal images and providing visual insights through interpretability techniques, it helps streamline clinical decision-making, reduces reliance on purely manual assessments, and enhances its potential for real-world application.

Author 1: Laberiano Andrade-Arenas

Author 2: Inooc Rubio Paucar

Author 3: Cesar Yactayo-Arias

Keywords: Classification; deep learning; Grad-CAM; medical imaging; ResNet-50

Paper 83: Case-Based Reasoning Model for Predicting the Malaria Cases

Abstract: Malaria remains a major public health issue in Ivory Coast, where the need for accurate and interpretable predictive models is critical for effective disease control. While most existing approaches prioritize predictive accuracy over interpretability, this study addresses the need for explainable models suitable for deployment in resource-limited public health settings. The study used a dataset covering multiple regions of Cˆote d’Ivoire over the period 2019–2023 and including climatic variables such as temperature, humidity, and precipitation. Seven conventional machine learning models (CatBoost, XGBR, RFR, DTR, SVM, KNN, and Linear Regression) were compared with three proposed Case-Based Reasoning variants (CBR 0, CBR 1, and CBR 2), which differ in their similarity-weighting strategies and correction constant. The results show that CBR 2 achieved the best predictive performance, with RMSE = 0.081, MAE = 0.057, and R2 = 72.40%, followed by the Random Forest Regressor. A Wilcoxon signed-rank test confirmed a statistical significant difference of this permofance (W = 18029, p = 1.02 × 10−20). Beyond predictive accuracy, qualitative criteria including explain-ability, transparency, and adaptability were evaluated, further highlighting the superiority of CBR 2 over conventional black-box models. These findings highlight the potential of Case-Based Reasoning for epidemiological forecasting and decision support in malaria control.

Author 1: Konan N’gatta Aimé Kouassi

Author 2: Koffi Kouakou Ive Arsene

Author 3: Gooré Bi Tra

Keywords: Malaria prediction; case-based reasoning; machine learning; epidemiological forecasting; explainable artificial intelligence

Paper 84: A Hybrid Semantic-Statistical Feature Fusion Framework for Bilingual Text Classification on Multilingual Big Data Corpora

Abstract: As the volume of scientific information is growing exponentially in several languages, there is a need for practical and scalable bilingual classification systems for large aligned scientific text corpora. To address this challenge, this study makes two key contributions - first, a large-scale bilingual English–Hindi aligned arXiv scientific text corpus, providing a structured resource for multilingual scientific text analytics and classification research is constructed. Second, the study proposes a bilingual scientific text classification framework and performs a rigorous experimental evaluation using three strong multilingual transformer models, namely Mini Language Model-12 Layers (MiniLM-L12), Multilingual Bidirectional Encoder Representations from Transformers (mBERT), and Cross-lingual Language Model-Robustly Optimized Bidirectional Encoder Representations from Transformers Pretraining Approach (XLM-RoBERTa), on the developed English–Hindi aligned arXiv big data corpus. English and Hindi summaries are categorized independently to investigate the performance trade-offs in each language. The proposed hybrid MiniLM-L12 + Multi-Layer Perceptron (MLP) architecture enhance the classification capability through the integration of statistical feature design with contextual sentence embeddings. Empirical analysis indicates that the proposed bilingual classification framework consistently outperforms the baseline transformer-only models, achieving a higher accuracy of 95.56% and weighted F1-score of 95.31%, while maintaining computational efficiency. The findings emphasize the effectiveness of hybrid representation learning for bilingual big data corpora and provide practical insights for scalable multilingual scholarly text analytics.

Author 1: Kavitha M

Author 2: Purohit Shrinivasacharya

Author 3: Y S Nijagunarya

Keywords: Text classification (bilingual); aligned big data corpora; multilingual transformers; mini language model-12 layers; multilingual bidirectional encoder representations from transformers; cross-lingual language model-robustly optimized bidirectional; encoder representations from transformers pretraining approach; multi-layer perceptron; hybrid deep learning

Paper 85: Security from Design, Bridging Model-Driven Architecture and DevSecOps Using Zynerator

Abstract: We propose an extension to Zynerator, a Model-Driven Architecture framework for automated microservice generation, that embeds DevSecOps principles directly at the modeling stage through semantic decorators. These decorators enable the automated synthesis of secure back-end and front-end components together with operational artifacts, including authentication and authorization modules, audit trails, monitoring dashboards, and DevSecOps pipelines covering SAST, DAST, testing, and deployment. The approach addresses a key limitation of the original Zynerator framework, namely the absence of explicit DevSecOps integration, and supports a security-by-design methodology that reduces reliance on specialized DevSecOps expertise. Through a detailed e-commerce case study and empirical evaluation against manual development and existing Model-Driven Architecture tools, we show that the enhanced framework reduces development effort, strengthens security posture, and accelerates DevSecOps adoption. These findings indicate that DevSecOps-aware model-driven engineering offers a viable pathway toward secure, auto-mated software delivery.

Author 1: Younes Zouani

Author 2: Mohamed Lachgar

Author 3: Youssef Harrati

Author 4: Mohamed Hanine

Author 5: Sulieman S. Alshuhri

Author 6: Amal Alomran

Keywords: DevOps; DevSecOps; MDA; IT security; code automation; semantic modeling; LLM

Paper 86: Mitigating Data Migration Risks in the Cloud via GA-Optimized Hybrid Cryptography Mechanisms

Abstract: Cloud computing has become one of the leading paradigms for bulk data storage and retrieval. However, ensuring data security remains a critical challenge, particularly during transmission when data is most vulnerable. Ensuring data security during transmission remains a critical challenge. The traditional algorithms focus mainly on reducing execution time and using static keys, which can create patterns in the ciphertext that an attacker can exploit. This study introduces an optimized ECC-AES-GA algorithm, which uses the Genetic Algorithm (GA) for generating an optimized parameter for ECC and bulk data is encrypted using the AES-256 algorithm. This algorithm provides security against man-in-the-middle, eavesdropping, replay, brute-force, impersonation, and forward secrecy attacks. The algorithm is also tested with the state-of-the-art algorithms and provides better results in terms of encryption/decryption and security parameters. Experimental analysis are performed based on avalanche tests, entropy level, throughput, execution time, and 10-times-run tests. Furthermore, the algorithm passed all necessary NIST STS tests, which confirms its cryptographic randomness and reliability. It offers security and efficiency, which improves computational overhead, thereby strengthening secure data migration in cloud environments.

Author 1: Anjali Dhaman

Author 2: Ugrasen Suman

Keywords: ECC-AES-GA; cloud data migration; optimization; GA (Genetic Algorithm); ECC (Elliptic Curve Cryptography); AES- 256; NIST statistical tests

Paper 87: An Explainable Hybrid AI Framework for Climate-Driven Environmental Health Risk Prediction in Agro-Ecosystems

Abstract: Climate change and environmental variability increasingly affect human health, particularly in agroecosystems exposed to fluctuating air quality and climatic conditions. Al-though recent advances in artificial intelligence have improved environmental risk prediction, many existing approaches operate as black-box systems and provide limited support for transparent decision-making and actionable interventions. This study presents an explainable hybrid artificial intelligence framework for climate-driven environmental health risk prediction. The proposed framework integrates environmental monitoring data, including Air Quality Index (AQI), temperature, and humidity measurements collected from publicly available environmental sources, with ensemble machine learning models (Random Forest, XGBoost, and LightGBM), SHAP-based explainability, and a Retrieval-Augmented Generation (RAG) module. Unlike conventional prediction systems, the proposed approach combines interpretable machine learning with evidence-grounded recommendation generation to enhance both transparency and practical usability. Experimental results indicate that XGBoost achieves the highest predictive performance, reaching an accuracy of 0.88 and an AUC of 0.91. SHAP analysis identifies AQI as the most influential factor affecting environmental health risk, followed by temperature and humidity. Furthermore, the RAG module was evaluated in terms of retrieval relevance and recommendation consistency, demonstrating its ability to generate context-aware recommendations supported by scientific knowledge sources. The proposed framework extends existing environmental health prediction approaches by jointly integrating predictive modeling, explainability, and knowledge-driven reasoning within a unified decision-support system. The results highlight its potential for supporting proactive environmental health management and climate-resilient decision-making in agroecosystems.

Author 1: Fatima-Zahra Alaoui

Author 2: Laila El Jiani

Author 3: Sanaa El Filali

Author 4: Rachida Ait Abdelouahid

Author 5: Zouheir Banou

Keywords: Explainable Artificial Intelligence; environmental health; climate change; XGBoost; SHAP; Retrieval-Augmented Generation; agroecosystems; machine learning

Paper 88: Behavior-Aware Access Control for IoT Networks Using Lightweight Machine Learning at the Gateway Level

Abstract: The growing amount of heterogeneous devices with scarce resources is compromising the security of the Internet of Things (IoT), as they are more likely to adapt to a fixed and identity-based access control. Conventional security systems tend to assume that once a device is authenticated, the network may be exposed to credential theft, firmware, and insider abuse. In this study a behavior-sensitive access control solution is presented, which integrates lightweight supervised Machine Learning (ML) on the IoT gateway to provide dynamic authorization. Unlike the traditional passive intrusion detection models, the proposed framework uses a Supervised Random Forest model to process real-time statistical feature summaries in terms of mean, standard deviation, and sparsity of the IoT telemetry data. The method converts the output of anomaly detection directly into access (full, restricted or blocked) levels. The system was implemented on a Flask-based gateway and tested with ToN-IoT benchmark dataset. The results of the experiments show an anomaly-class recall of 0.9986 (99.86%) with 91,169 correctly detected attack and 125 false negatives among the 91,294 attack instances, for a security-oriented Zero Trust profile. As an example, when rounded to two decimal places, this value is 1.00, but the unrounded value is reported so as not to suggest 100% detection. The enforcement layer focuses on reducing risk and removes or filters out requests that were determined to be malicious or unauthorized in the scenarios. The architecture is designed to provide low latency through feature extraction and inference on the edge, which provides data privacy because telemetry processing is locally done without relying on the cloud.

Author 1: Yaseen Alduwayl

Author 2: Abdullah Alessa

Author 3: Mounir Frikha

Keywords: Behavioral-aware access control; edge intelligence; gateway-based security; IoT security; intrusion detection; lightweight machine learning; random forest

Paper 89: A New Drilling Rate of Penetration Prediction Model by Particle Swarm Optimization and Gradient Boosting Regression

Abstract: The oil and gas industry continuously evolves to enhance operational efficiency and productivity while minimizing costs and environmental impact. Among the critical aspects of oil and gas operations, drilling efficiency is a key factor in accessing underground hydrocarbon reservoirs. Traditional machine learning models and current regression models have shown limitations in accurately modelling the rate of penetration due to the high nonlinearity of data. This project focuses on the rate of penetration prediction for drilling optimization. This study proposed a new drilling rate of penetration prediction model with the embedding of particle swarm optimization in a gradient boosting regression method. A solution representation of the particle is introduced as a hyperparameter strategy to explore the optimal parameter for predicting drilling datasets. Extensive experiments were carried out using two splitting strategies, One for All and All for One, across the drilling wells. The proposed GBR+PSO hybrid model achieved a mean absolute error of 1.205, representing a reduction of approximately 89.68% compared to the best-performing baseline model, K-Nearest Neighbor with the One for All splitting strategy, which achieved a mean absolute error of 11.68. The hybrid solution could enhance drilling ROP predictions, advancing the drilling rate of penetration strategy. It has the potential to support the development of autonomous drilling optimization, thus contributing to more efficient, reliable, and cost-effective drilling rate of penetration strategies in future drilling operations.

Author 1: Faris Aiman Jamaluddin

Author 2: Marina Yusoff

Author 3: Diva Kurnianingtyas

Author 4: Mohamad Taufik Mohd Salledud-din

Keywords: Drilling; gradient boosting regression; machine learning; rate of penetration; particle swarm optimization

Paper 90: An Experimental Evaluation of Deep Learning Networks for Automated Breast Cancer Detection

Abstract: Breast cancer remains one of the leading causes of cancer-related mortality among women worldwide, where early and accurate diagnosis plays a vital role in improving survival rates. Recent advancements in deep learning have demonstrated significant potential in automating the analysis of medical images for cancer detection. This study presents a comprehensive comparative analysis of convolutional neural network (CNN)–based deep learning models for breast cancer classification using ultra-sound and mammography images. Multiple architectures, including a baseline CNN, AlexNet, DenseNet, ResNet50, ResNet101, VGG16, VGG19, and MobileNetV3, were evaluated to classify breast lesions as benign or malignant. Experimental results reveal notable performance differences across imaging modalities. For ultrasound images, AlexNet achieved the highest accuracy of 89%, while DenseNet and the baseline CNN achieved 88% and 85%, respectively. In contrast, mammography-based classification yielded significantly higher performance, with the baseline CNN outperforming deeper architectures in terms of accuracy and F1-score, achieving 97%. The findings demonstrate that model complexity does not necessarily guarantee superior performance and that properly designed shallow CNNs can effectively outperform deeper networks on high-quality mammographic data. This study highlights the potential of deep learning–based computer-aided diagnosis systems to support radiologists in the early detection of breast cancer.

Author 1: Partha Chakraborty

Author 2: Umme Aiman Jannat

Keywords: Breast cancer; deep learning; convolutional neural networks; mammography; ultrasound; medical image classification

Paper 91: Optimization of Service Function Chain Placement in Cloud-Fog-Edge Networks

Abstract: There is an explosion in IoT devices, 5G technology, and MECs, that results in increasing demands on effective and scalable network services management. Service function chaining, defined as the sequence of functions in VNFs on a path, is one of the core principles behind the NFV architecture design. SFC allocation to the heterogeneous clouds–fogs–edges network is an NP-hard problem characterized by mutually conflicting goals, such as latency minimization, energy and cost reduction, and resource maximization. In this study, an in-depth comparison study is carried out on three population-based optimization algorithms for solving the placement problem of SFCs using Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), and Grey Wolf Optimization (GWO) on three cases: (1) VNF deployment cost/QoE optimization in a 5G hybrid cloud with 12 nodes and weighting factor γ=0.4; (2) SFC graph matching on MEC-NFV networks with a 100-node physical network, 20 VNFs, and equal utilization weights α=β=γ=1/3; and (3) multi-instance SFC mapping on Fog-to-Cloud (F2C) IoT environment with a 5-VNF chain across 5 nodes. These three algorithms have been evaluated under identical conditions: 10 independent runs, 200 iterations, 20–30 agents. Results demonstrate that GWO achieves the best VNF deployment objective (W =73.92, a 15.1% improvement over the BGWO baseline), PSO achieves the highest resource utilization (52.9%) in MEC-NFV placement, and both PSO and GWO reduce F2C end-to-end latency by 25% compared to the ILP reference (12 vs. 16 units at three instances), while all three algorithms reduce latency by approximately 80% relative to cloud-only deployment. PSO emerges as the most consistently high-performing algorithm across all three scenarios.

Author 1: Chandrapal Singh Dangi

Author 2: Sanjay Sharma

Keywords: Service function chaining; Virtual Network Functions; Network Function Virtualization; Particle Swarm Optimization; Ant Colony Optimization; Grey Wolf Optimization; mobile edge computing; Fog-to-Cloud; 5g networks; resource allocation

Paper 92: 2I-CSO: A Novel Intelligent and Interoperable Cat Swarm Optimizer Approach to Optimal Cluster Head Selection and Self-Termination Search

Abstract: Wireless sensor networks (WSNs), fundamental building block of IoT, are subject to several constraints because of the finite non-rechargeable energy resources available in the nodes. The selection of Cluster Head (CH) plays a critical role in determining the energy balance in a network. Conventional methods such as the LEACH algorithm choose CH randomly with a probability mechanism that might lead to choosing weak nodes as CHs and thereby fail prematurely. The biologically -inspired optimization methods, such as PSO and CSO, help to enhance CH selection using a global approach. However, these methods suffer from the following three major shortcomings: 1) random switching between the exploration and exploitation stages, 2) lack of intelligence during the formation of clusters, and 3) growing exponentially complex search space of CHs.This study proposes an Intelligent and Interoperable Cat Swarm Optimizer (2I-CSO), a protocol designed to address these limitations simultaneously. 2I-CSO also introduces an interoperable configuration mechanism based on LEACH’s hierarchical architecture, where the Base Station maintains a centralized energy configuration table shared with Cluster Heads and member nodes, ensuring network-wide parameter consistency and enabling the intelligent stopping mechanism. Experiments conducted on five well-known TSPLIB test cases and WSN simulations demonstrate that 2I-CSO outperforms individual metaheuristics. Simulation results on a custom web-based platform further show that 2I-CSO achieves faster convergence, lower computational cost, and competitive network lifetime compared to standard CSO and the Emperor Penguin Optimizer (EPO). To the best of our knowledge, the proposed intelligent stopping condition is the first introduced for bio-inspired WSN clustering protocols.

Author 1: Oumaima Hassan

Author 2: Mohammed Essaid Riffi

Keywords: Wireless sensor networks; Cat Swarm Optimization; Cluster Head Selection; energy efficiency; bio-inspired optimization; configuration space reduction; intelligent stopping condition; Emperor Penguin Optimizer; network lifetime

Paper 93: Emotion-Aware Serendipity Recommendation from Textual Reviews: BERT-Based Emotional Clustering

Abstract: This study proposes an emotion-aware serendipity recommendation framework based on textual reviews. Six BERT-based binary classifiers are trained to detect surprise, curiosity, trust, nostalgia, frustration, and enchantment from user reviews. The predicted emotion probabilities are used to build stable review-level emotional vectors. Two complementary contributions are introduced. Contribution 1 develops an emotion-aware reranking strategy that combines calibrated emotional information with relevance, novelty, and unexpectedness. Contribution 2 introduces EC-SereRank, an emotional-combination-based reranking model that preserves co-activated emotions within reviews and groups them using K-medoids into interpretable affective meta-groups. Experiments on Amazon Product Reviews show that affective information improves serendipity-oriented recommendation. The ablation study confirms that relevance and unexpectedness alone are not sufficient, while the integration of emotional compatibility improves recommendation quality. EC-SereRank further shows that emotional combinations provide an interpretable complementary signal for affective serendipity recommendation.

Author 1: Mariam Benayad

Author 2: Ahmed Zellou

Keywords: Recommender systems; emotion-aware recommendation; textual reviews; BERT; K-medoids; K-means

Paper 94: QuantumGuard: A Post-Quantum Resilient Deception-Driven Framework for Proactive Threat Hunting and Cognitive Honeynet Orchestration in 6G-Enabled Cyber-Physical Systems

Abstract: The convergence of 6G ultra-reliable low-latency communications, massive cyber-physical actuation, and the looming threat of cryptographically relevant quantum computers exposes a fundamentally new attack surface that is poorly addressed by reactive intrusion-detection paradigms. This study presents QuantumGuard, a proactive cybersecurity framework that inverts the conventional defender posture by integrating cognitive honeynets, post-quantum-secured intelligence channels, and reinforcement-learning-driven deception adaptation across 6G-enabled cyber-physical environments. The principal contribution is the architectural integration of four previously disjoint capabilities—a cognitive honeynet orchestrator, a reinforcement-learning deception policy engine operating under partial observability, a transformer-based MITRE ATT&CK attribution network, and a post-quantum federated intelligence bus se-cured with CRYSTALS-Kyber and CRYSTALS-Dilithium—into a single closed-loop control architecture for 6G cyber-physical systems. To assess feasibility, we report a preliminary evaluation on the CICAPT-IIoT-2024 and Edge-IIoTset benchmark datasets, complemented by a 320-endpoint 6G slice testbed configured, as described in Section IV. Initial results indicate an attacker engagement-retention rate of approximately 96.42 per cent, MITRE ATT&CK technique-attribution accuracy of approximately 91.8 per cent, a 78.6 per cent reduction in median attacker dwell time relative to passive honeypot baselines, and a deception-induced production-traffic overhead of 1.7 per cent. The PQ-FIB sustains a 14.2 ms median post-quantum handshake latency at slice scale. We position these numbers as preliminary evidence of operational viability under the evaluated threat model rather than as fully characterized performance bounds; a follow-up empirical study, planned for a separate publication, will extend the evaluation to deception-aware adversaries and sustained-attack stress conditions.

Author 1: Daifallah Zaid Alotaibe

Keywords: Cyber deception; cognitive honeynet; post-quantum cryptography; reinforcement learning; 6G security; cyber-physical systems; threat attribution

Paper 95: Predicting Histological Progression in Primary Biliary Cirrhosis Using Advanced Machine Learning Techniques

Abstract: Cirrhosis is considered one of the most serious liver diseases worldwide, closely related to excessive alcohol consumption and inadequate eating habits, factors that progressively deteriorate people’s health. In this context, the present research aimed to develop and validate a predictive model based on Machine Learning (ML), specifically using the Random Forest (RF) algorithm, to determine the histological stage (from 1 to 4) in patients with primary biliary cirrhosis. For the development of the study, the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology was applied, which includes the stages of business understanding, data understanding, data preparation, modeling, and evaluation. The results obtained showed that the model presents better performance in the more advanced stages of the disease. The area under the curve (AUC) increased from 0.612 in Stage 1 to 0.874 in Stage 4, reflecting a notable improvement in its discriminative capacity. Similarly, metrics such as sensitivity, precision, and F1-score showed an upward trend, reaching their highest values in Stage 4. In this sense, the proposed model represents a complementary diagnostic support tool, since it allows estimating the histological stage through the analysis of clinical data, contributing to medical decision-making without relying exclusively on invasive procedures.

Author 1: Laberiano Andrade-Arenas

Author 2: Cesar Yactayo-Arias

Author 3: Inoc Rubio Paucar

Keywords: Clinical decision support; histological stage prediction; primary biliary cirrhosis; random forest

Paper 96: Deep Learning for Canonical Reconstruction of Deformable Objects from Depth Images in Robotic Vision and SLAM-Aware Perception

Abstract: Reconstructing the canonical pose of non-rigid objects from arbitrary depth observations is an important problem in robotic vision, particularly for systems that must perceive, track, and interact with deformable objects in dynamic environments. In robotics and SLAM-related perception, depth cameras are widely used to support object recognition, spatial understanding, scene mapping, and motion analysis. However, non-rigid deformation remains challenging because the same object may appear in significantly different poses, making reliable object-level representation and tracking difficult. In this study, we present a deep learning approach for reconstructing the default canonical pose of non-rigid objects from single depth images. The proposed model combines short-range and long-range feature extraction with the original depth input to capture both local geometric details and global structural information. By transforming arbitrary posed observations into a consistent canonical representation, the method supports more stable shape understanding, pose normalization, and object-level perception for robotic systems operating in real-world environments. This is particularly relevant to robotic vision tasks involving human motion analysis, deformable-object tracking, manipulation, and semantic mapping. The model is trained on synthetic human datasets and evaluated on synthetic human, real human, and animal datasets. Experimental results demonstrate improved retrieval accuracy compared with existing methods, showing that the proposed approach can generalize across different non-rigid categories and sensing conditions. These findings highlight the potential of canonical pose reconstruction as a useful component for intelligent robotic perception, depth-based scene interpretation, and SLAM-aware systems that require robust understanding of deformable objects.

Author 1: Fahd Alhamazani

Keywords: Depth image; computer vision; feature representation; sensor fusion; object-level SLAM; human–robot interaction

Paper 97: Accelerating Blockchain Consensus: A Parallel Mining Approach for High-Throughput, Low-Latency Networks

Abstract: Problem: Blockchain consensus remains con-strained by slow transaction verification, redundant mining effort, high communication overhead, and increased confirmation latency as the number of nodes grows. Objective: This study proposes NCABS, a controlled parallel mining consensus approach intended to improve blockchain throughput, latency, transaction commit rate, and resource utilization. Methodology: NCABS combines verifiable head-miner selection, transaction-validation delegation, parallel block generation, and a 75%confirmation threshold. The framework is evaluated against Practical Byzantine Fault Tolerance (PBFT) through ten repeated experiments over node densities from 40 to 140, using throughput, latency, communication load, transaction commit rate, scalability, and responsiveness as performance metrics. Results: The plotted results show that NCABS achieves higher throughput and transaction commit rates than PBFT, reduces aggregate communication messages by approximately 45%, and lowers latency at 40 nodes from about 310 ms for PBFT to about 195 ms for NCABS. In the transaction-commit-rate experiment, NCABS reaches approximately 3200 transactions per second, compared with about 2200 transactions per second for PBFT. Conclusion: The findings indicate that controlled parallel mining can reduce redundant computation and communication while improving blockchain responsiveness. The current evaluation is performance-oriented; adversarial testing and broader empirical comparison with additional consensus baselines are identified as future work.

Author 1: Sohail Jabbar

Keywords: Blockchain; consensus algorithm; parallel mining; PBFT; high transaction rate; low-latency

Paper 98: Explainable Neural Network Prediction of Post-COVID-19 Depression via Monte Carlo Simulation

Abstract: The COVID-19 pandemic has driven a substantial rise in depression, notably in Argentina’s post-quarantine period, motivating the need for predictive tools to support timely mental health interventions. This study uses a Feedforward Neural Network (FNN) and Monte Carlo simulations to predict depression scores from key socio-economic and psychologicalvariables—anxiety state, economic income, and education—benchmarked against SVR, GRU, Linear Regression, Decision Tree, and Random Forest. The FNN achieved the best overall performance (MAE = 4.72, RMSE = 6.32, R2 = 0.64; cross-validated R2 = 0.593 ± 0.048), while Linear Regression attained the highest R2 (0.693), suggesting partly linear relationships among predictors. Monte Carlo simulations showed that higher anxiety increased predicted depression, while higher income and education reduced it, underscoring the value of targeted anxiety-reduction and economic-support interventions in post-pandemic mental health policy.

Author 1: Siham AKIL

Author 2: Sara SEKKATE

Author 3: Abdellah ADIB

Keywords: Depression prediction; Feedforward Neural Net-work; Monte Carlo simulations; explainable artificial intelligence; post-COVID-19 mental health; socio-economic factors; SHAP analysis

Paper 99: Context-Aware Sentiment Analysis of E-Commerce Reviews Using a BERT-CNN-BiLSTM Hybrid Model

Abstract: User-generated product reviews are an essential source of information in e-commerce; nevertheless, the huge volume and varying quality of review texts make extracting insights difficult. The conventional approach to sentiment classification is limited in terms of recognizing contextual and aspect-oriented sentiment clues in the text. This study proposes a hybrid architecture that uses contextual sentiment clues for e-commerce reviews sentiment classification. The experiments are conducted using FABSA dataset which contains reviews annotated in terms of multiple aspects–sentiments pairs. Every review is expanded in terms of aspect-oriented sentiment classification samples. This makes it possible to learn fine-grained polarities of sentiments associated with specific aspects of product reviews. To tackle the problem of class imbalance in the data, the experiments employ the method of stratified oversampling in combination with the use of class-weighted cross-entropy loss function. The empirical results show that the improved hybrid BERT–CNN–BiLSTM model achieves 92.20% validation accuracy, 92.25% weighted F1-score, and 86.36% macro F1-score. The most noticeable progress has been made in the case of neutral class, where the F1-score has increased by 17.1 percentage points showing the improvements in minority-class recognition and decrease of class imbalance. The architecture-level ablation study shows the superior performance of the BERT–CNN–BiLSTM model compared to its simplified versions based on BERT, BERT–CNN, and BERT–BiLSTM architectures. A contextual comparison with reported FABSA baselines suggests competitive performance, although this comparison should not be interpreted as a strict leaderboard result because the baselines were not reproduced under identical experimental settings. Overall, the findings demonstrate that combining contextual transformer representations with local convolutional features and bidirectional sequential modeling can improve class-balanced sentiment classification in aspect-expanded e-commerce review data.

Author 1: Mahmud Reza Mahim

Author 2: Fahad Bin Z Islam

Author 3: Md Saklain Mahmud

Author 4: Md. Imtiyaz Hasan

Author 5: Sifat Rahman Ahona

Keywords: Sentiment analysis; e-commerce reviews; BERT; CNN; BiLSTM

Paper 100: Advanced and Classical Selection Methods in Genetic Algorithms: A Comprehensive Comparative Analysis

Abstract: Selection mechanisms critically influence the convergence behavior and solution quality of Genetic Algorithms (GAs). This study presents a rigorous empirical comparison of six selection methods: three classical methods—Random Selection, Roulette Wheel Selection (RWS), and Tournament Selection (TS)—and three adaptive methods: Fitness-Distance Balance (FDB), Dynamic FDB (dFDB), and Functional Weight-based Selection (FW). Experiments were conducted across 23 classical benchmark functions (F1–F23) and 10 CEC2019 functions (cec01–cec10), with each configuration executed 30 times using consistent GA parameters. Performance was assessed using Best, Mean, Median, and Standard Deviation, with statistical significance determined by the Wilcoxon rank-sum test (α = 0.05). The results reveal that TS consistently achieved the best or statistically equivalent performance in 30 out of 33 functions, outperforming both classical and adaptive alternatives. Notably, RWS showed surprising competitiveness, outperforming adaptive methods such as FDB and dFDB in several scenarios. While dFDB and FW improved over static FDB, they failed to consistently outperform TS. These findings confirm TS as a robust default choice for diverse optimization landscapes and provide new empirical evidence regarding the limited practical advantage of current adaptive strategies within GAs. This study contributes the first controlled GA-based evaluation of adaptive selection mechanisms on both classical and CEC2019 benchmarks, offering insights for practitioners designing efficient evolutionary systems. Limitations related to fixed GA settings, function diversity, and adaptive method complexity are acknowledged, and future work is suggested to explore hybrid and problem-aware selection strategies.

Author 1: Husam S. Mashaqbeh

Author 2: Putra Sumari

Author 3: Hamza A. Mashagba

Author 4: Mohammed Hashem Almourish

Author 5: Azlan Abd Aziz

Author 6: Lara A. Al-Mashagba

Author 7: Wael Waheed Alqassas

Keywords: Genetic algorithm; selection mechanisms; FDB; tournament; roulette wheel; evolutionary computation; optimization

Paper 101: Combatting Phishing Attacks: Leveraging Machine Learning for Real-Time Detection in Penetration Testing

Abstract: Phishing attacks continue to pose a significant threat to individuals and organizations, driven by the increasing sophistication of cybercriminal techniques and the rapid expansion of digital services. Traditional detection approaches, such as blacklist-based and rule-based systems, are often ineffective against newly generated or obfuscated phishing URLs. This study proposes a machine learning (ML)-based framework intended for integration within penetration testing environments. The approach leverages multiple supervised learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), and XGBoost, trained and evaluated using the PhiUSIIL Phishing URL Dataset, a large-scale benchmark dataset containing phishing and legitimate URL samples. A comprehensive preprocessing pipeline and feature engineering strategy are employed to enhance model performance. Experimental results demonstrate exceptionally high detection accuracy, with RF and XGBoost achieving near-perfect classification performance across key evaluation metrics, including accuracy, precision, recall, F1-score, and ROC-AUC. The proposed system is further designed for real-time deployment, enabling integration into penetration testing workflows for proactive security assessment. Despite promising results, limitations related to dataset characteristics and real-world generalization are acknowledged. Overall, this research highlights the effectiveness and practical applicability of ML-based approaches in strengthening phishing detection and advancing modern cybersecurity defences.

Author 1: Ashwag Alotaibi

Author 2: Mounir Frikha

Keywords: Phishing detection; machine learning; real-time detection; penetration testing; URL analysis

Paper 102: A Model for Refactoring Monolithic Applications to Microservices Using Domain-Driven Design: A Case Study on IoT Irrigation Systems

Abstract: The increasing complexity of Internet of Things (IoT) applications has exposed the limitations of monolithic software architectures in addressing scalability, flexibility, and real-time processing requirements. Although microservice architectures offer a promising alternative, identifying optimal service boundaries remains a significant challenge, often resulting in excessive inter-service communication and degraded system performance when poorly defined. This study proposes a quantitative model for refactoring monolithic applications into microservices by integrating Domain-Driven Design (DDD) principles with measurable metrics, including service size, coupling, and scalability. The model systematically identifies optimal service boundaries through a structured evaluation framework. The proposed approach is validated using a case study of an IoT-based irrigation management system. Experimental results show a reduction in inter-service communication overhead and improved modularity and scalability compared to baseline decomposition approaches. The findings demonstrate that combining DDD concepts with quantitative analysis provides an effective and practical solution for guiding microservice migration in complex IoT environments. The average coupling score across the refactored system was recorded at 20.4%, which satisfies the theoretical requirement of remaining below 30% and aligns with empirical observations from successful microservice decompositions.

Author 1: Munezero Immaculee Joselyne

Author 2: Ngenzi Alexander

Author 3: Hitimana Eric

Author 4: Ipinnimo Oluwafemi

Keywords: Microservice; refactoring model; IoT irrigation system; DDD

Paper 103: A Multi-Criteria Decision-Making Model for ERP Selection in Moroccan SMEs Using Fuzzy AHP

Abstract: Today, selecting an Enterprise Resource Planning (ERP) system has become a strategic decision for Small and Medium-sized Enterprises (SMEs) seeking to improve their operational efficiency, modernize their information systems, and strengthen their competitiveness in an increasingly digitalized environment. In this context, identifying the most relevant ERP selection criteria is essential to support decision-makers in choosing the solution that best fits their organizational and strategic needs. However, previous studies have generally examined ERP selection criteria from a broad perspective, without prioritizing their relative importance according to the specific context of SMEs. The objective of this study is therefore to identify and prioritize the ERP selection criteria that are most critical for Moroccan SMEs, as well as to propose a decision-making model that supports ERP evaluation and selection. To achieve this objective, a mixed-methods approach was adopted. Semi-structured interviews with ERP professionals and practitioners were conducted to identify the most relevant criteria in the Moroccan SME context. In addition, the Fuzzy Analytic Hierarchy Process (Fuzzy AHP) method was used to evaluate and weight these criteria under conditions of uncertainty and subjective judgment. The results reveal that functional scope is the most critical criterion in ERP selection, followed by ease of use and total cost. Technical criteria such as compatibility and security are considered moderately important, whereas scalability appears to have less influence on the selection decision. Based on these findings, a decision-making model is proposed to help SMEs evaluate ERP alternatives more effectively and reduce the risks associated with ERP selection. This research contributes to the existing ERP literature by providing a structured and prioritized analysis of ERP selection criteria specifically adapted to Moroccan SMEs. It also offers ERP consultants and decision-makers a practical framework for assessing ERP solutions and improving the quality and reliability of ERP selection decisions.

Author 1: Yassine Zouhair

Author 2: Younous Elmrini

Keywords: ERP selection; Fuzzy AHP; Moroccan SMEs; multi-criteria decision-making; information systems; decision framework; digital transformation