ABSTRACT. Machine learning techniques, particularly deep learning using the long short-term memory (LSTM) algorithm, have been increasingly employed for stock market prediction. Although these methods have been applied to developed markets, only a limited number of studies have examined emerging markets. This paper introduced a deep learning model that utilized LSTM to predict emerging stock markets and investigated the impact of exogenous variables on these predictions. The three metrics, root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE), were used to evaluate the model’s performance accuracy. The datasets were extracted from four emerging stock market indices: Qatar, Saudi Arabia, Brazil, and China. The experimental results highlighted the role of external factors in improving prediction accuracy. The highest percentage of stock market prediction improvement was 33.33% of MAPE in Saudi Arabia, while the lowest MAPE was 13.16% in Brazil. The lowest improvement percentage was 2.27% based on the RMSE of the Qatar stock market, whereas the highest RMSE improvement was 23.04% in China. The MAE improvement percentages were 13.63%, 13.96%, 26.67%, and 31.90% for Qatar, Brazil, China, and Saudi Arabia, respectively.

ABSTRACT. With the sophistication of cyber attacks, traditional “fix” rule-based Intrusion Detection and Prevention (IDPS) systems, such as Snort, are limited. While Snort is able to detect known attack signatures, it's not very adaptable to zero day and emerging threats and doesn't offer automated realtime response features. To overcome these drawbacks, the present study proposes a hybrid real-time IDPS which combines Snort 3 with machine learning to enhance the accuracy of detection, reliability of verification, and automated mitigation. The proposed system utilizes a Random Forest classifier to do the classification on the CIC-IDS2017 dataset for the features that will be extracted from the network traffic being captured in real time by TShark. Suspected flows are confirmed by means of Snort signature rules, and confirmed threats are automatically blocked by iptables. The ELK Stack can also be utilized to visualize security events and alerts to monitor in real-time. It was implemented in a virtual testbed comprising of Kali Linux, Metasploitable 3, and Ubuntu. Experimental results revealed 98% accuracy, low false positive and real-time detection/response to SSH brute force and DDoS attacks. The overarching contribution is the ability to combine all the different anomaly detection, verification, monitoring and automated response components within a single real-time IDPS workflow, using machine learning techniques.

ABSTRACT. Efficient feeding in poultry farms is essential to reduce feed waste, improve farm production and sustainability. This study is an Automated Feed Level Detection Feeding System Using Convolutional Neural Network to detect feed pellets in chicken feeding plates. The system captures image of the chicken feed plates and process to the trained CNN model to count and to be able to classify as sufficient or insufficient. The Experimental validation consists of 20 trials under varying initial weights and scheduled amounts. Over those trials the dispensing actuator produced a mean dispensing error of −0.07 g, indicating minimal in dispensing, with a mean absolute error (MAE) 1.17 g, and a root mean square error (RMSE) 1.25 g, demonstrating a consistent dispensing accuracy. The CNN feed-level predictor showed a mean predicting error of +0.41 g, mean absolute error 1.03 g, and RMSE 1.16 g. These results indicate that the system effectively monitors feed availability and dispenses appropriate amounts while minimizing waste. The study shows that integrating computer vision with automated dispensing of feeds provides a reliable, cost-effective, and scalable solution for intelligent poultry feeding management.

ABSTRACT. In this capstone project, a Smart Code Lab was designed as a real-time desktop IDE that addresses inefficiencies in programming laboratory teaching through an offline, LAN-based solution, overcoming the problems of manual code grading, poor instructor visibility, and unreliable internet access. The system combines a student-facing IDE with syntax highlighting, local code execution, and automated static analysis of Java, Python, and C++, as well as a central instructor-facing dashboard to monitor real-time coding activity, compilation failures, and submission. An automated evaluation module provides immediate feedback evaluating code against instructor-defined criteria of correctness, robustness, efficiency, and maintainability. Functional testing confirmed all test cases on session management, live monitoring, and rule-checking, and controlled single-user performance benchmarks demonstrated fast startup times, low memory usage, and quick LAN discovery, though full-classroom network stability under concurrent load remains theoretical and is identified as future work. User acceptance testing among students and faculty produced high satisfaction ratings across functionality, reliability, and usability. Smart Code Lab is a C#-based standalone Windows application, which makes it dependable in connectivity-constrained environments and lowers the burden on the instructor through timely intervention, improving programming education in resource-constrained academic laboratories.

ABSTRACT. An adaptive ensemble approach for real-time hazard perception and collision avoidance in autonomous vehicles (AVs) is proposed to address the critical challenge of reliable detection under dynamic and adverse driving conditions, including urban streets, highways, fog, and low-visibility environments. A novel model selection method is introduced, in which three state-of-the-art deep learning models (YOLO, Faster R-CNN, and SSD) are dynamically combined within a unified multi-sensor fusion pipeline using camera, LiDAR, and radar data. The framework is evaluated on multiple datasets, including KITTI, Waymo Open Dataset, nuScenes, and BDD100K, as well as on synthetic environments generated by CARLA, NVIDIA DriveSim, and AirSim. Experimental results indicate that the YOLO model is outperformed by Faster R-CNN in terms of accuracy, with 92% accuracy and a response time of 150 ms, while YOLO achieves 95% accuracy and a response time of 50 ms. The SSD model is characterized by a response time of 100 ms and an accuracy of 93%, thereby providing a balance between accuracy and speed. It is demonstrated that the robustness of detection is further improved by approximately 8–12% under adverse conditions when multi-sensor fusion is employed, compared to vision-only baselines. The proposed adaptive ensemble, which is based on real-time risk assessment and time-to-collision estimation, is shown to achieve up to 7% improvement in F1-score compared to single-model deployments and prior results on the KITTI and Waymo datasets. The feasibility of adaptive model selection and multi-sensor fusion is thus validated as an effective and scalable approach for enhancing AV safety frameworks.

ABSTRACT. Systemic lupus erythematosus (SLE) flares are difficult to anticipate, yet early prediction is essential to prevent organ damage and avoid unnecessary treatment escalation. In practice, data are fragmented across hospitals and cannot be easily centralized due to privacy and governance constraints. This work addresses 30-day SLE flare prediction under such constraints using heterogeneous clinical, laboratory, and wearable time-series data distributed across multiple sites. Conventional statistical models and single-site deep learning approaches rely on limited modalities, assume pooled datasets, and often fail to capture long-range temporal structure and cross-patient relationships. We propose FedFlares, a privacy-preserving multimodal federated learning framework that combines FEDformer-based temporal modeling with a graph neural network to encode patient similarity while keeping all patient records local. Experiments on a multicenter cohort of 730 patients and 12{,}040 visits with 1{,}740 documented flares show that FedFlares achieves an AUROC of 0.93, an AUPRC of 0.73, and an F1-score of 0.74, outperforming centralized Transformer baselines and standard federated methods. Modality and ablation analyses confirm that multimodal fusion, long-horizon temporal modeling, and graph-based contextualization each contribute to these gains. Overall, FedFlares provides accurate, well-calibrated flare risk estimates in a realistic federated setting, supporting its potential integration into distributed clinical decision-support workflows for SLE management.

ABSTRACT. This paper presents an ensemble machine learning approach to predict the success of first-time Social Work Licensure Examination takers at Western Mindanao State University (WMSU), addressing the absence of systematic, data-driven early-warning systems for at-risk students. A dataset of 140 first-time examinees was collected across 2019–2025, covering demographic, academic, socioeconomic, psychological, and behavioral variables. The dataset was augmented to 1,110 records using the Synthetic Data Vault (SDV) Gaussian Copula Synthesizer. Six base models and six ensemble configurations such as bagging (Random Forest), boosting (Gradient Boosting), and stacking variants were trained and evaluated via 10-fold cross-validation using RMSE as the primary metric. SHAP (Shapley Additive Explanations) provided post-hoc feature importance analysis. Results identified study hours (SHAP = 0.376), GWA (0.286), and test anxiety (0.262) as the three most influential predictors. On the hold-out test set (n = 30), the Augmented Base Ridge model achieved the best performance (RMSE = 1.6628, MAE = 1.3677), followed closely by the Stacking Neural-Ridge-Ridge Final (RMSE = 1.6673). Data augmentation improved neural network-based models (Base Neural Network: RMSE from 11.80 to 2.19), while traditional models remained stable. The top seven models on unseen data were all augmented-data models, confirming that synthetic augmentation strengthens generalization. The findings are operationalized in a web-based decision-support system for faculty to identify and support at-risk students before the licensure examination.

ABSTRACT. Accurate heart disease prediction (HDP) remains a critical challenge in medical informatics for reducing global health risks. While existing machine learning (ML) and deep learning (DL) models have been applied to HDP, there is a persistent need to enhance predictive accuracy. This study proposes a stacking ensemble deep learning (SE-DL) model to improve HDP performance. The SE-DL framework integrates three pre-trained DL architectures—Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU) with two meta-learners: Logistic Regression (LR) and Support Vector Machine (SVM). The model’s effectiveness was evaluated using accuracy, recall, precision, and F-1 score metrics. Experimental results demonstrated that the SE-DL model significantly outperformed individual DL and ML algorithms. Notably, the proposed model achieved superior results, 96.59% accuracy, 96.7% precision, and 96.59% F1-score, when applied to datasets without feature selection compared to several feature selection methods. This indicated its robust internal feature extraction capabilities.

ABSTRACT. ZoneShield is an experimental fire protection device designed to be a small-scale four-zone device that can help test whether combining the logic of fire protection systems used in industry and inexpensive sensors and microscale robotic nozzles can work together. This hydraulic system includes a pre-charged tank line, a pressure-balancing jockey pump, two main pumps, branch lines controlled by solenoid valves, IR-based infrared flame detectors, and an Arduino/ESP8266-based control panel. The Automation Studio software was used to verify the pump and valve sequence of operation before hydrostatic testing to optimize pressure limits, pump startup time delays, and failure and non-recovery scenarios without setting fire. Afterwards, five laboratory experiments were conducted to assess the system's performance. These experiments included the following: discharge of one zone, guided discharge using a servo motor, simultaneous discharge of two zones, restart condition, and pressure reduction without fires. The findings revealed that the fire detection time was below 0.5 seconds, with an average of roughly 0.3 seconds. Tests on one zone found that the water would take between 0.5 and 0.7 seconds to reach the point of target, while ethanol tank fires of five centimeters in diameter would be extinguished in roughly two to three seconds. The two zones would take about four seconds to extinguish. Pressure data showed that the pressure was able to increase from roughly 1.2 bar to almost 3 bar when operating only one zone, and maintained a consistent and operable value of 2.5 bar for two zones operated simultaneously. These findings reveal potential to incorporate backup pressure, phase pump operations, and direct nozzles into an effective laboratory model, although there are currently no indications that it can be applied in the field.

ABSTRACT. —The manual analysis of X-ray images for the diagnosis of bone fractures is a complex and time-consuming process that is prone to human error. The majority of current deep learning-based analyses are binary fracture detection (fracture vs no fracture), making them less useful for fracture treatment planning, where the specific fracture type is needed. In this study, a deep learning-based system for multi-class classification of 12 different types of bone fractures that commonly occur in orthopedic practice is proposed. A dataset of 3,449 X-ray images was collected from various sources (Kaggle and a hospital in Sana’a city, Yemen) using a curated one. These have been expert-annotated into 12 fracture categories. Transfer learning was performed using ImageNet-pretrained weights for DenseNet-121. The data set was divided into three: training (70%), validation (10%), and test (20%). Training used the Adam optimizer (lr=0.001), categorical crossentropy loss, batch size 32, and class weights to compensate for imbalance. The performance of test accuracy for DenseNet-121 is 86.29% (95% CI: 84.1–88.5%), macro precision of 0.854, macro recall of 0.847, and macro F1-score of 0.849. The macro-average AUC was 0.94. The results showed that the fractures were perfectly classified (F1=1.000), and the lowest performance (F1=0.712) was obtained for oblique fractures because of the similarity in appearance to hairline fractures. Deep learning can be used to classify fractures in multiple classes. Multi-class fracture classification using deep learning is feasible. Furthermore, achieves clinically relevant performance.

ABSTRACT. This paper aims to assess the roles and responsibilities of the Chief Information Officer (CIO) in the Yemeni telecom sector. It addressed the vagueness surrounding the CIO role, which has led to uncertainty regarding expectations, practices, and outcomes associated with this position. Empirically, the paper investigates the CIO role through the perspectives of professional information technology (IT) managers and IT employees in the Yemeni telecom sector. The participants were selected based on their expert knowledge of the people of IT who are working in the headquarters of the three Yemeni telecom companies, Sabafon, YOU, and Yemen Mobile. To achieve the research objective, the purposive sampling technique was used, and a questionnaire-based survey was developed as a data collection tool to elicit responses to the research questions. A total of 113 questionnaires were distributed manually and online to the three different telecommunications operators in Yemen. Only 90 valid questionnaires were received, resulting in a response rate of 79.78%. The findings highlight the multifaceted contributions of effective CIOs within the Yemeni telecommunications sector, identifying them as technology providers, strategic supporters, business thinkers, innovation drivers, integration advisors, and relationship managers.

ABSTRACT. This study examines how digital banking services contribute to customer loyalty in the banking sector of Aden, Yemen, and whether customer experience explains this relationship. The study responds to limited evidence on post-adoption loyalty in underrepresented and service-constrained banking contexts. Data were collected from 323 retail banking customers and analyzed using confirmatory factor analysis and structural equation modeling with an estimator suitable for ordinal survey responses. Digital banking services were represented by speed and responsiveness, security and trust, ease of use, personalization, and channel integration. Customer experience was represented by cognitive perception, digital interaction, emotional impression, perceived quality, and perceived value. The results show that stronger digital banking services improve customer experience and strengthen loyalty. Customer experience also plays an important mediating role, showing that digital channels support loyalty when they create secure, smooth, integrated, and value-enhancing interactions. The study concludes that banks in Aden should manage digital banking not only as a delivery channel, but also as an experience-building capability that supports retention, relationship continuity, and customer recommendation.

ABSTRACT. Traditional dental diagnosis from radiographic images is often subjective, prone to human error from fatigue, and inaccessible for individuals in resource-limited settings, leading to diagnostic inconsistencies and delayed treatment. This study addresses this gap by developing "TootHack" a web-based deep learning system designed to provide an accessible, objective, and accurate diagnostic aid for both dental professionals and the general public. Centered on a Mask R-CNN model with a ResNet-101 backbone. A dataset of 13,932 panoramic radiographs was sourced from public repositories and underwent significant manual data cleaning and re-annotation of segmentation masks to ensure data integrity. With manual validation by licensed dentists, achieving an average: Precision of 80.63%, Sensitivity of 90.81%, F1-Score of 85.37%, and an Accuracy of 89.71%. Furthermore, initial usability testing with selected participants yielded a System Usability Scale (SUS) average score of 74; Excellent in usability.

ABSTRACT. Online product reviews heavily influence consumer purchasing decisions. However, the rise in the number of fake reviews poses great risk to the reputation of the platforms. In this paper, we introduce ReviewAI, an integrated system which combines natural language processing with a hybrid machine learning classifier system for the detection of deceptive texts. The pipeline integrates three different classification models: Support Vector Machine (SVM), Random Forest, and DistilBERT. In the experiment using independent data from Amazon and Shopee, our hybrid ML ensemble model employs the weighted soft voting technique to achieve out-of-domain test accuracy of 94.0%, surpassing other models (SVM: 87.0%, Random Forest: 93.0%). For deployment purposes, ReviewAI operates as a Google Chrome browser extension and web app through Django API, which performs online inference of review sets in less than 3 seconds.

ABSTRACT. Generating high-quality candidates in high-dimensional, discrete sequence spaces remains a fundamental challenge for generative artificial intelligence. While traditional Reinforcement Learning (RL) methods like REINFORCE are effective for objective-driven optimization, they are frequently prone to mode collapse and reward hacking, where agents converge to simplistic, redundant solutions to exploit sparse reward signals. In this study, we propose a robust generative framework that integrates a Stack-augmented Recurrent Neural Network (StackRNN) with Proximal Policy Optimization (PPO) to enforce trustworthy and stable policy updates. By employing a clipped surrogate objective, our approach constrains the update step within a defined trust region, preventing destructive parameter shifts during fine-tuning. Comparative benchmarks against the REINFORCE baseline demonstrate that the PPO-driven agent significantly enhances distributional stability and manifold exploration efficiency. Principal Component Analysis (PCA) reveals that our framework achieves a more expansive and continuous coverage of the structural manifold, successfully identifying complex, high-novelty architectures that the baseline fails to capture. These results establish PPO as a superior strategy for ensuring the reliability and diversity of generative agents, offering a scalable pathway for AI-driven discovery in complex search spaces.

ABSTRACT. Conventional genomic signal processing (GSP) pipelines often fail to differentiate position-specific codon constraints and omit structural proteomic descriptors such as Electron-Ion Interaction Potential (EIIP) physical residue mappings. In this paper, we introduce a unified prosodic-genomic framework that fuses genomic syntax, triplet cadence, and proteomic physical properties to evaluate structural protein hotspots via integrated cross-layer EIIP narrowband analysis. We establish robust Bandpass Network (BPN) baselines for genomic periodicity tracking alongside a dedicated proteomic EIIP filter track for localized hotspot classification. Benchmarked against the AF009614 dataset, our proposed framework achieves superior classification accuracy at strict operational percentile constraints ($81^{\text{st}}$, $82^{\text{nd}}$, and $85^{\text{th}}$ percentiles). Integrating explicit metric templates, continuous phase-slip tracking ($\mathit{WaveSync}$), and an enhanced mutation weighting module ($\mathrm{RBI}_+$) significantly minimizes false-positive annotations while improving localized boundary detection accuracy at verified binding interfaces.

ABSTRACT. This study examines how electronic customer relationship management practices shape customer loyalty in Yemeni retail banking and explains the mediating role of customer trust. Drawing on Commitment-Trust Theory, the paper treats E-CRM as a higher-order relationship capability reflected by customization, security and privacy, payment options, problem handling, feedback and online contact, and help content. Data were collected from retail banking customers who had used electronic banking or bank-managed digital interaction channels and were analyzed using covariance-based structural equation modeling. The measurement results supported the reliability and validity of the proposed model, including the multidimensional structure of E-CRM. The structural findings show that E-CRM improves customer trust and customer loyalty, and that trust is the strongest pathway through which E-CRM contributes to loyalty. The mediation results indicate that customer trust carries the largest share of the relationship between E-CRM and loyalty, while E-CRM also retains a smaller direct contribution. The study contributes a validated empirical model for understanding how digital relationship practices generate loyalty in an underrepresented banking market. Practically, the findings suggest that banks in Yemen should design E-CRM around trust-sensitive priorities, including secure interaction, transparent communication, responsive problem handling, reliable follow-up, and accessible digital support.

ABSTRACT. This study aims to address the issues faced by dog owners in terms of maintaining regular feeding time and achieving accurate and individualized portion control in compliance with veterinary standards, which otherwise leads to underfeeding, overfeeding, and behavioral issues in dogs. Most conventional and basic forms of automated dog feeding systems lack identification and real-time monitoring systems. Therefore, to address the issues faced by dog owners, the proposed PAW FEEDS is an advanced Internet of Things-based automated dog feeding system, which incorporates RFID identification systems, load cell systems, and real-time scheduling through a dedicated React Native mobile application for the optimal nutrition of dogs. The proposed system incorporates a two-microcontroller configuration for optimal performance. The proposed system incorporates an ESP32 microcontroller for all hardware-related tasks, scheduling, servo motor control, RFID, and load cell. The second ESP32-CAM module is used for real-time video streaming. The portion control is dynamically calculated through the mobile application using standard veterinary calculations, i.e., Resting Energy Requirement (RER) and Maintenance Energy Requirement (MER), and real-time data is synchronized through Firebase Cloud Services. The proposed system has been tested for reliability and accuracy, and the results show 95% success in RFID-based pet identification, 97% accuracy in food dispensing, and 0.45 seconds average latency in cloud communication. The usability of the proposed system has also been tested among 35 evaluators, and the results show a high degree of satisfaction, achieving an average satisfaction score of 4.42 out of 5. The proposed system is a significant contribution to the field of IoT-based smart homes for the automation of pet feeding systems, as it incorporates veterinary calculations and RFID access along with real-time video streaming.

ABSTRACT. Enterprise resource planning systems are increasingly adopted by international non-governmental organizations to strengthen accountability, coordination, and service delivery. However, implementation outcomes remain uneven because ERP success depends not only on technology, but also on how organizations engage stakeholders throughout the change process. Drawing on stakeholder theory, socio-technical thinking, and information systems success literature, this study examines how stakeholder involvement relates to ERP implementation success in humanitarian INGOs operating in Aden, Yemen. A quantitative survey of 150 ERP users examined stakeholder analysis and representation, participation in ERP lifecycle activities, communication and information sharing, and readiness and support for change. ERP implementation success was assessed through system quality, information quality, service quality, user satisfaction, and organizational benefits. The findings indicate that all stakeholder-involvement dimensions are positively associated with ERP implementation success, with readiness and support for change showing the strongest role. The study positions stakeholder involvement as an integrated organizational capability that supports ERP outcomes in conflict-affected humanitarian settings and offers practical guidance for improving ERP implementation through representation, participation, communication, training, and change support. The findings should be interpreted as associative evidence from Aden-based INGOs rather than causal evidence generalizable to all humanitarian organizations.

ABSTRACT. Digital transformation has become an important organizational issue for INGOs seeking to improve coordination, responsiveness, and service quality in demanding operating environments. This study examines how digital transformation relates to service delivery performance in INGOs operating in Aden and whether organizational learning explains part of this relationship. A cross-sectional survey of 184 employees was analyzed using PLS-SEM. The findings indicate that digital transformation strengthens organizational learning and contributes to service delivery performance both directly and indirectly through organizational learning. The results suggest that digital tools are more likely to improve service outcomes when they support knowledge sharing, experience documentation, reflection, and adaptive follow-up across organizational units. The study contributes context-specific evidence from INGOs in Aden and offers practical implications for development-oriented organizations seeking to align digital systems with learning routines and service improvement.

ABSTRACT. As the Vietnamese electric vehicle (EV) market transitions toward mainstream adoption, understanding generational differences in consumer behavior is crucial. This study investigates the determinants of EV purchase intention, specifically comparing Generation Z and Generation X. Primary data were collected from 502 prospective consumers at automobile dealerships using a structured questionnaire. Applying Partial Least Squares Structural Equation Modeling (PLS-SEM) and Multi-Group Analysis (MGA), the study evaluated the effects of governmental assistance, EV technology, charging infrastructure, environmental concern, and perceived price.The findings reveal distinct generational decision-making pathways. For Gen X, purchase intention is significantly driven by all five factors, with charging infrastructure and environmental concern being the most critical. Conversely, Gen Z is predominantly motivated by cutting-edge EV technology, while completely disregarding perceived price. Furthermore, MGA confirms that the influence of EV technology is significantly stronger for Gen Z than for Gen X. These results suggest that automobile manufacturers and policymakers must abandon "one-size-fits-all" approaches. Strategies should be tailored to position EVs as advanced smart devices for Gen Z, while addressing practical, infrastructural, and economic anxieties to ensure adoption among Gen X.

ABSTRACT. This study investigates the strategic potential of implementing drone-integrated last-mile delivery services within dense urban environments, using Can Tho City as a focal empirical setting. Grounded in the Technology Acceptance Model (TAM) and behavioral theory, the proposed conceptual framework elucidates five fundamental relationships encompassing perceived ease of use (PEOU), delivery speed, perceived usefulness (PU), positive attitude, and behavioral intention (BI). Empirical data derived from 390 valid survey responses were subjected to rigorous statistical analysis, including Cronbach’s alpha, convergent and discriminant validity assessments, and Structural Equation Modeling (SEM) via AMOS. The empirical results provide robust support for all five hypothesized paths. Notably, perceived usefulness emerged as the most significant predictor of behavioral intention (beta = 0.454), followed by the impact of speed on consumer attitude (beta = 0.450) and the influence of PEOU on perceived usefulness (beta = 0.419). The structural model exhibited satisfactory fit indices (CFI = 0.915, GFI = 0.900, RMSEA = 0.078). This research contributes a parsimonious, yet theoretically sound model tailored to the Vietnamese urban landscape, outlining the technology acceptance behavior of young urban consumers. From a managerial perspective, the findings underscore that logistics firms must prioritize the demonstration of utilitarian benefits, specifically expedited delivery timelines while streamlining interface navigation to foster favorable consumer attitudes and consolidate long-term adoption intentions.

ABSTRACT. The Rivest–Shamir–Adleman (RSA) cryptosystem is fundamental to secure data transmission, yet its computational intensity presents significant performance challenges, particularly when processing large files. This paper presents a parallel implementation of RSA using OpenMP to leverage multi‑core processors. A large plaintext is partitioned into independent blocks, each encrypted or decrypted concurrently using static scheduling with thread affinity. Experimental evaluation on an Intel Core i5‑8365U (4 physical cores, 8 logical threads, 16 GB RAM) using key sizes of 1024, 2048, and 4096 bits and message sizes up to 50 MB achieves a peak decryption speedup of 20.23× and a total execution speedup of 19.71× for the combined encryption and decryption processes (4096‑bit key, 50 MB, 18 threads). Encryption‑only operations peak at 4.61× (4096‑bit key, 30 MB, 6 threads). Notably, even under the heaviest workload (18 threads, 4096‑bit key, 50 MB), the system maintains an efficiency of 1.10, confirming excellent scalability and efficient parallel resource utilization. Our implementation incorporates CRT decryption, thread affinity, adaptive static chunk scheduling, and per‑thread buffer reuse, significantly outperforming prior OpenMP‑based works [7]– [9]. The parallelization strategy indicates that larger key sizes, traditionally associated with increased computational cost, actually enable better parallel efficiency due to the increased per‑block computational work that amortizes parallel overhead [5], [23].

ABSTRACT. Generative Artificial Intelligence (GenAI) is reshap- ing education by enabling personalized learning, automated content generation, and intelligent instructional support.However, adoption of GenAI in education remains fragmented, with limited frameworks that connect tool-level integration to broader peda- gogical and institutional transformation. To address the gap, this paper proposes conceptualized GenAI adoption framework for education from integration and augmentation to transformation (GIFT). The GIFT captures how educational systems evolve from basic AI-assisted tasks (integration), to human–AI collabora- tive learning (augmentation), and ultimately to fully redesigned pedagogical and institutional practices. GIFT is defined across five dimensions (i.e., AI role, teacher role, assessment design, student agency, and institutional support). GIFT dimensions enable a multi-level understanding of how GenAI reshapes teaching, learning, and governance processes across stages. By providing a unified and interpretable model, GIFT supports classification of GenAI adoption, guides institutional strategy, and offers a foundation for future validation and policy development. The framework contributes to advancing evidence-informed and pedagogically meaningful integration of GenAI in education.

ABSTRACT. The rapid growth of healthcare big data has increased the need for scalable machine learning (ML) systems that can handle large medical imaging tasks. In radiology, imaging methods like MRI and CT produce data streams that require a lot of computing power and memory. Therefore, High-Performance Computing (HPC) configurations despite major advancements in deep learning models, documentation on system-level scalability validation, deployment testing, and reproducibility is still lacking. Therefore, this review identifies common gaps and challenges, highlights an imbalance between algorithm development and comprehensive system-level HPC evaluation. Specifically, the reviewed studies exhibit poor reporting of hardware configurations and GPU-hour usage, a lack of strong and weak scaling analyses, limited evaluation of latency and throughput under realistic clinical conditions, and insufficient assessments of cost and energy consumption. This research examines a comprehensive literature review focused on system-level scalability and deployment within radiology. The quantitative findings reveal that 36.7% of the reviewed studies concentrate on single-GPU or multi-GPU training approaches. However, only 26.7% address inference performance when deployed at scale, and 20% of the research discusses federated and distributed methods, though these studies frequently omit important details about communication protocols and associated costs. The review highlights a critical need for establishing standardized approaches to reporting computational metrics, performance measures that are clinically meaningful, and documentation that allows for reproducible scalability results. Implementing these enhancements would help close the existing gap between achieving high model accuracy and successfully deploying these systems in actual healthcare environments.

ABSTRACT. The proposed research examines the challenges of Information Technology Service Management (ITSM) for a Yemeni governmental organization (GO) following the pilot implementation of the Information Infrastructure Library (ITIL) framework. This research tends to be a model for integrating artificial intelligence (AI) technologies with ITIL4 to improve ITSM. An approach based on hybrid methods, combining applied and quantitative methods, was used in a simulated environment. AI features, including natural language processing (NLP) based chatbots, automated routing, and predictive analytics, were employed to measure key performance indicators (KPIs). The research outcome represented a lower mean time to repair (MTTR), while the repeat request percentage (RRP) decreased by 51.10%. This study concludes that integrating AI with the ITIL 4 framework is essential for enhancing organizational governance. It transforms traditional service management into proactive and autonomous management.

ABSTRACT. Water distribution networks (WDNs) are a critical backbone of urbanservice infrastructure, where pipe integrity is to be sustained inseparably from public health and efficient utilization of resources. Undetected leakages place significant economic burdens and water quality risks, but reliable fault identification under actual field conditions is a truly unsolved issue. Rule-based detectors and conventional classifiers are operationally convenient but not very accurate, pressure-threshold methods only achieve 86.1\% and support vector machine (SVM) approaches stalling close to 90.4\%, especially under variable demand and sensor noise. To address this gap, an Internet-of-Things (IoT)-to-edge digital twin (DT) framework is proposed, wherein hydraulic residuals from a calibrated EPANET simulation are fused with edge-preprocessed sensor features for continuous leak identification. Experiments on the publicly available GraphLeak benchmark, using stratified 70/10/20 splits and 30-second sliding windows, yield a detection accuracy of 96.8\%, an F1-score of 95.9\%, an area under the receiver operating characteristic curve (AUROC) of 0.989, a top-1 localization accuracy of 84.9\%, a mean localization error of 88~m, and an edge inference latency of 18.6~ms per window. These results support the practical viability of physics-informed edge intelligence for smart water network monitoring.

ABSTRACT. Several Yemeni archipelagos suffer from chronic energy insecurity due to a costly and inefficient reliance on diesel power generation. Hybrid renewable energy systems (HRES) offer a sustainable, environmentally friendly solution for these remote island regions. This paper conducts a techno-economic optimization and evaluation of an HRES configuration—comprising photovoltaic (PV) panels, wind turbines (WT), hydro turbines (HT), and battery storage—tailored for Mayyun Island to provide a sustainable and cost-effective energy supply. HOMER Pro software was employed to simulate and evaluate three system designs based on net present cost (NPC), cost of energy (COE), and excess/deficit energy metrics. The optimal design yielded a COE of US$0.0616/kWh and an NPC of US$119,079, which are significantly lower than current diesel-based generation costs, while maintaining unmet load and excess electricity at 0.075% and 28%, respectively. These results confirm that integrating multiple renewable energy streams enhances grid stability, minimizes battery storage requirements, and mitigates resource intermittency. Ultimately, these findings provide critical insights for stakeholders and policymakers driving renewable energy deployment in remote sectors.

ABSTRACT. Forecasting gold prices is a challenging task due to nonlinear market dynamics, macroeconomic uncertainty, and rapidly changing investor sentiment. This paper proposes TFA-Net, a Temporal-Fusion Attention Network that integrates Long Short-Term Memory (LSTM), Random Forest, and an adaptive attention mechanism for gold price forecasting. The framework combines historical gold prices, macroeconomic indicators, and sentiment scores extracted from financial tweets using FinBERT. The LSTM component captures temporal dependencies in sequential market data, while the Random Forest component models nonlinear feature interactions. An attention-based fusion layer dynamically weights the contribution of each information stream according to market conditions. Experiments were conducted using daily gold market data from January 2024 to October 2025 with a next-day forecasting horizon. Results demonstrate that TFA-Net outperforms standalone LSTM and Random Forest models, achieving the lowest forecasting error with a MAPE of 1.2%. The findings confirm that hybrid architecture combined with sentiment-aware modeling significantly improve forecasting performance and robustness in volatile financial environments.

ABSTRACT. Motor function prediction in children with cerebral palsy can be strengthened by combining neuromuscular signals and speech motor cues. Multimodal learning enables joint modeling of these complementary biomarkers for objective assessment. However, EMG and speech recordings vary widely across children and sessions and often contain noise, which reduces the reliability of unimodal predictors. Clinical use also benefits from calibrated confidence along with accurate outcomes. Conventional methods typically use handcrafted EMG or speech features with classical learners, and fusion is often performed by feature concatenation or score averaging. Such fusion does not learn cross-modal alignment and is vulnerable to missing channels and noisy audio. NeuroVox-CP is proposed as a self-supervised EMG--speech fusion framework using cross-attention to learn shared representations and an uncertainty-aware head for reliable predictions. The model targets GMFCS severity grading and GMFM-66 functional score estimation under subject-wise evaluation.The study uses 120 pediatric cerebral palsy participants after quality control, with two sessions per child separated by 7--14 days. Standardized upper-limb motor tasks and controlled speech tasks (sustained /a/ and /pa-ta-ka/) are recorded with therapist-rated GMFCS and GMFM-66 labels.NeuroVox-CP attains GMFCS Accuracy 0.87, Macro-F1 0.85, and AUROC 0.95, and predicts GMFM-66 with MAE 3.8, RMSE 5.4, and $R^2$ 0.84. Robustness is maintained at SNR 10 dB (Macro-F1 0.83) and with two missing EMG channels (Macro-F1 0.82), with improved calibration (ECE 0.028). Overall, self-supervised cross-modal fusion improves accuracy and reliability for pediatric motor outcome prediction. The framework supports confidence-aware functional assessment and longitudinal monitoring in clinical settings.

ABSTRACT. Accurate CP severity scoring can be strengthened by jointly modeling neuromuscular activity and speech articulation. However, unimodal learning and weak fusion often fail to capture cross-modal correspondence, leading to unstable severity estimates. Conventional EMG-only or speech-only deep models, as well as simple concatenation/gated fusion, typically learn modality-specific embeddings without explicit alignment. This work proposes \emph{BioFusionNet}, a contrastive EMG--speech alignment framework with attention-guided fusion to learn a shared severity-relevant latent space and improve prediction consistency. Experiments use public datasets capturing complementary impairment cues: NinaPro DB1 (27 subjects, 52{,}200 EMG samples at 2~kHz) and TORGO (15 subjects, $\sim$18~h dysarthric speech at 16~kHz). BioFusionNet achieves the best performance (Accuracy = 0.993, Macro-F1 = 0.982, MCC = 0.972, $\kappa$ = 0.968) and the lowest errors (MAE = 0.028, RMSE = 0.054), while ablations confirm the critical role of cross-modal fusion (Accuracy: 0.993 $\rightarrow$ 0.974 when removed). The results indicate that contrastive alignment with adaptive fusion yields more reliable multimodal severity scoring.

ABSTRACT. Today's supply chains are increasingly complex and data-rich, with disruptions, demand fluctuations, and disjointed data making it difficult to maintain transparency and resilience. This paper introduces a Business Intelligence-Driven Architecture (BIDA) that enhances supply chain visibility, responsiveness, and resilience by leveraging real-time analytics, predictive monitoring, and adaptive decision support. The suggested architecture comprises of four interconnected parts: data acquisition, data integration and storage, business intelligence and analytics, and decision support and resilience. The layers create a robust data-to-decision pipeline that converts heterogeneous operational data into actionable insights to inform effective supply chain management. The efficiency of the proposed architecture was analyzed through simulation and, compared with a traditional logistics system. Experimental results show that the improvements are significant in various performance metrics. The proposed BIDA framework further improved the Supply Chain Visibility Index (SCVI) by 40.3% and the Resilience Index (RI) by 39.7%. Moreover, the disruption recovery time was shortened by 42.8%, data latency dropped by 31%, and cost efficiency increased by 22%. These results show increased agility in operations, quicker disruption response, and better decision-making. The proposed framework also offers the possibility of a shift from reactive to proactive supply chain management, using predictive analysis and early warning systems. Future work will be directed towards implementation and integration with emerging intelligent technologies in the real world.

ABSTRACT. Enhancing low-resolution facial images remains an important challenge in computer vision, especially in surveillance and face recognition applications operating under limited visual quality conditions. In this paper, the effectiveness of lightweight super-resolution models in improving facial images by optimizing training rather than increasing architectural complexity is investigated. The experiments were conducted on the CelebA dataset using a scaling factor of ×4. No modifications were made to the network topology; instead, the original Super-Resolution Convolutional Neural Network (SRCNN) was enhanced through preprocessing, adaptive learning-rate scheduling, and extended training epochs. For comparison, the Efficient Sub-Pixel Convolutional Neural Network (ESPCN) and Very Deep Super-Resolution (VDSR) models were evaluated under the same experimental conditions. Reconstruction quality was assessed using the Peak Signal-to-Noise Ratio (PSNR), the Structural Similarity Index Measure (SSIM), RGB-based evaluation, and the Learned Perceptual Image Patch Similarity (LPIPS) metric. The optimized SRCNN model achieved better reconstruction performance than both bicubic interpolation and a previously reported SRCNN-based implementation. VDSR achieved the highest PSNR and SSIM values of 28.56 dB and 0.8751, respectively, and demonstrated strong perceptual reconstruction performance according to LPIPS evaluation. These findings indicate that lightweight super-resolution models can provide an effective balance between reconstruction quality and computational efficiency when supported by appropriate training optimization.

ABSTRACT. E-financial systems, such as digital payment systems, e-wallets, and merchant applications, have radically altered the aspect of financial services in the world by making them more accessible, efficient, and more essential in daily life. Nevertheless, with the fast evolution of online financial services, it has also made it easier to fall victim to digital financial threats, which includes phishing, vishing, ATM skimming, QR code scams, fake loan placements, and online employment scams. As much as the efforts of countries in promoting financial inclusion have greatly increased access to digital financial services, the levels of digital financial literacy that are demanded to safely utilize such services are highly lopsided among the groups of users. The paper discusses the risk awareness and vulnerability of women schoolteachers across multiple districts of Kerala, India, to digital financial risks using a quantitative cross-sectional research design. Convenience sampling of educational institutions was used to come up with a sample of 112 female schoolteachers. To determine the knowledge of participants about digital financial services, awareness of frequent cyber-financial threats, and exposure to digital financial threats, a structured self-designed questionnaire was given. To measure vulnerability along five dimensions, the Digital Financial Vulnerability Index (DFVI) was created, and inferential statistics analyses were conducted to establish the predictors of vulnerability. The results of this research provide empirical data to a poorly researched field of digital financial literacy, which provides information on the risk perception and knowledge gap among women educators. The findings have practical implications for government structures, financial regulators, and community-based stakeholders in the development of specific digital financial literacy programs. Finally, building the digital financial resilience of women schoolteachers could create safer digital financial practices at homes, in schools, and in communities—building more inclusive and resilient digital financial ecosystems.

ABSTRACT. Recidivism is one of the most criminal issues that significantly affects the security of the community. It also affects the stability and capacity of governments, particularly in conflict zones such as Yemen. Many research in the literature were recently proposed to predict the likelihood of a person's relapse into criminal behavior after the release (recidivism). However, there are different shortcomings related to the previous methods especially their performance and applications. In this paper, we propose a hybrid machine learning framework to predict the possibility of the occurrence of recidivism. The proposed framework has four contributions. The first is to build the in-house data set, of 3,110 records from the Marib Correctional Facility covering 2021–2025, to be used to evaluate the performance of the framework. The second is to utilize SMOTE technology with the XGBoost algorithm to address significant imbalances between classes. The third is SHAP analysis, which is used to interpret XGBoost decisions. Finally, the K-Means algorithm was used to distribute the prisoner into seven behavioral groups according to the international Risk-Need-Responsivity (RNR) model. The results showed that the proposed framework achieved significant performance. It achieved accuracy, precision, recall, F1-score, and ROC-AUC with 99.19%, 99.15%, 96.72%, 97.92%, and 0.9998, respectively. The SHAP analysis confirmed that criminal records and annual crime rates are the most important factors significantly mitigating demographic biases. The results showed that the proposed framework is promising and could be used as a practical roadmap for correctional institutes in resource-limited environments.

ABSTRACT. Advanced persistent threats (APTs) remain among the hardest cyberattacks to detect because they evolve over long dwell times, blend with legitimate traffic, and frequently bypass signature-based defenses. This paper proposed an explainable hybrid detection model for APT identification. Seven baseline learners were benchmarked, including XGBoost, CatBoost, multi-layer perceptron (MLP), DenseNet, autoencoder, TabNet, self-supervised learning, and 1D-CNN. The final architecture combines XGBoost and MLP in a hybrid stacking configuration and augments prediction with SHAP and LIME to support both global and local interpretability. The Model results show that the proposed model achieves 99.98% test accuracy, 99.81% recall, and 99.47% F1-score while maintaining a very low false-alarm rate.

ABSTRACT. The increasing sophistication of malware has highlighted the limitations of traditional signature-based detection methods, motivating the adoption of machine learning (ML) techniques. This review analyzes approximately 59 studies published between 2019 and 2026 that cover static, dynamic, and hybrid malware analysis approaches. The surveyed methods utilize features extracted from API calls, permissions, opcode sequences, and system behavior to improve detection accuracy. The review finds that static analysis offers efficiency and scalability. In contrast, dynamic analysis provides deeper behavioral insights but incurs higher computational costs. Hybrid approaches consistently achieve superior performance by combining the strengths of both techniques. This paper provides a comparative overview of existing machine learning-based malware detection methods, identifies current research trends and challenges, and highlights future directions for developing more robust and adaptive malware detection systems.

ABSTRACT. The huge revolution in communications and the internet has introduced modern techniques like the Internet of Things (IoT). IoT and the high dynamism of its devices face challenges such as routing, energy efficiency, and scalability. Traditional proactive routing protocols like OLSR are good in stable topologies but lack efficiency in dynamic environments. Therefore, enhancing OLSR becomes necessary to improve route selection in IoT networks. We propose a new approach that combines Ant Colony Optimization (ACO) and Deep Reinforcement Learning (DRL) to enhance the performance of the OLSR protocol in IoT networks. ACO is used for path exploration, while DRL handles route selection decisions. The hybrid model is implemented and evaluated using the NS-3 platform and the Python language. The simulation was carried out under varying node densities, traffic conditions, and data priorities. The performance was evaluated by the Packet Delivery Ratio (PDR), End-to-End Delay (E2ED), throughput, and routing overhead metrics. The proposed hybrid model achieves a Packet Delivery Ratio (PDR) of 48.64%, representing a noticeable improvement over standard OLSR (42.96%) and OLSR-ACO (32.45%). Furthermore, at a moderate density of 50 nodes, it successfully reduces end-to-end latency to 15.09 ms compared to 47.74 ms for standard OLSR, while maintaining stable throughput and a controlled control-plane overhead growth. These findings indicate that combining bio-inspired optimization with deep reinforcement learning provides an effective and scalable solution for intelligent routing in the next generation of IoT networks.

ABSTRACT. The continuously increasing usage of Internet of Things (IoT) technologies has led to greater vulnerabilities to various forms of cyberattacks. Therefore, it is very important to design a scalable and computationally efficient intrusion detection system (IDS) that is capable of working. This paper proposes an enhanced IDS framework that utilizes Linear Discriminant Analysis (LDA) for dimensionality reduction, thereby improving feature separability and reducing computational complexity. The study examines the trade-off between predictive performance and computational efficiency using four machine learning models: Random Forest (RF), KNearest Neighbors (KNN), Naive Bayes (NB), and eXtreme Gradient Boosting (XGBoost). Experiments were conducted on the UNSW-NB15 dataset across eight LDA feature extraction scenarios (from 2 to 9 features). The results show that Random Forest achieved the highest accuracy at 97.9%, while Naive Bayes demonstrated superior computational efficiency, with an execution time of 0.7 seconds. In addition, a comparison with recent related studies highlights the competitive efficiency and effectiveness of the proposed approach. Overall, the framework provides a balanced trade-off between accuracy and computational cost, making it suitable for real-time IoT security applications.

ABSTRACT. The integration of artificial intelligence (AI) in e-health systems is a potentially transformative innovation, but its adoption in fragile, post-conflict healthcare settings is severely underexplored. This study examined factors influencing AI adoption at Al-Gamhuria Modern General Hospital in Aden, Yemen, using a sequential explanatory mixed-methods approach (n = 143 surveys; n = 28 interviews). The exploratory factor analysis revealed that four dimensions, namely organisational, human, technical infrastructure and ethical-legal frameworks, had a strong structure, which explained 77.8% of the total variance. Multiple regression analysis showed that organisational support (β = 0.39, p < 0.001) and human factors (β = 0.32, p = 0.006) were the strongest predictors of adoption intention, whereas technical infrastructure was not significant (β = 0.15, p = 0.293). Four main themes arose from the qualitative thematic analysis: operational continuity as a prerequisite, workforce displacement concerns, legal-accountability gaps, and low-complexity offline tools as accessible entry points. We propose a contextual three-stage adoption framework that puts more emphasis on operational resilience than infrastructural perfection. The results contribute to TAM and UTAUT2 by proposing resilience as a sine qua non in volatile settings. Policy recommendations include interim national AI healthcare guidelines, hospital-level digital support units, and a phased roll-out of offline-capable diagnostic and decision-support tools.

ABSTRACT. Accurate fault detection of faults accurately continues to be an important issue within an industrial environment because of its complexity. There have been many attempts at research towards enhancing fault detection accuracy, especially within the field of semiconductor manufacturing. Therefore, this paper proposes a hybrid XGBoost–Random Forest model to address this research gap and advance the current state of the art. The Kaggle-sourced UCI-SECOM dataset was utilized, which presents several challenges, including missing values and highly imbalanced classes. To prepare this dataset, a holistic methodology was introduced to preprocess the data accurately and deal with the problem of unbalanced classes. This process involves the use of median imputation to deal with missing values as well as feature selection, which is aimed at eliminating redundant features as well as using the Synthetic Minority Over-sampling Technique (SMOTE). According to the experimental results obtained from this study, it has been found that the model accuracy was improved to 99.49%. When dealing with semiconductor manufacturing processes, early fault detection saves time and resources.

ABSTRACT. Customer reviews on online platforms have become a vital source of insight into organizational performance. Delivery companies increasingly rely on customer feedback to identify areas for improvement. Sentiment analysis is a computational technique used to identify and analyze the emotions and attitudes individuals express toward specific entities, products, services, or applications. These textual sentiments are commonly classified into three classes: positive, negative, and neutral. The purpose of this study is to enhance the performance of sentiment analysis for Arabic reviews related to delivery applications in Yemen by using auto-labeling with large language models (LLMs) and fine-tuning BERT-Based models, including AraBERT, MARBERT, and CAMeLBERT. Additionally, the study addresses the challenge of class imbalance in the dataset, which can significantly affect model performance, particularly for minority classes. The dataset comprises approximately 6,967 customer reviews extracted from five Yemeni delivery applications: Tawseel, Tasaheel, Wagbat, Etlobni, and Talka. A total of 6,508 reviews with consistent labels assigned by both LLMs were established and used to train and evaluate the models. Two training strategies were implemented and compared: Standard Training and Training with Imbalanced Data Handling. The experiments revealed that MARBERT achieved the highest performance, with a weighted-average F1-score of 98.56% and a macro-average F1-score of 95.16%.

ABSTRACT. Association Rule Mining (ARM) is a foundational paradigm for discovering hidden co-occurrence patterns within large-scale datasets. However, the rapid evolution of the big data era imposes severe computational constraints on traditional ARM methodologies such as Apriori and ECLAT, particularly regarding scalability, memory utilization, and realtime adaptability in high-velocity, dynamic environments. This paper presents a systematic and critical evaluation of recent architectural and algorithmic advances in ARM, specifically focusing on literature spanning 2023 to 2026. The synthesis prioritizes four critical dimensions: incremental mining frameworks, distributed and parallel architectures, domainspecific applications, and statistical interest-driven metrics. By evaluating the inherent trade-offs between execution throughput and structural memory footprints, this review delineates the computational bottlenecks of current state-of-the-art models. The findings reveal that while parallel frameworks successfully mitigate velocity challenges, they consistently suffer from structural memory explosions when scaling to high-dimensional datasets. Finally, this study conceptualizes a comprehensive architectural roadmap and highlights key open research directions, serving as a benchmark for designing next-generation, context-aware, and scalable ARM topologies

ABSTRACT. The Bab El-Mandeb Strait is a strategically important maritime chokepoint that requires reliable real-time situational awareness. Traditional cloud-based surveillance systems face communication latency, scalability limitations, and vulnerability to single points of failure, motivating the adoption of edge intelligence. This paper presents EdgeGuard, a simulation-driven edge AI framework for maritime threat detection based on adaptive multi-sensor fusion. The framework employs a CNN–LSTM architecture enhanced with attention-based fusion to improve robustness under varying sensing conditions. The system was evaluated in a simulated 12-node edge deployment scenario representing the Bab El-Mandeb region. Experimental results achieved 96.0% accuracy and a Macro F1-score of 96.1%, while the adaptive fusion mechanism improved Macro F1 by 5.3% under simulated adversarial conditions compared with a static fusion baseline. All experiments were conducted in simulation environments, and real-world deployment remains future work.

ABSTRACT. Abstract—The use of machine learning techniques for predicting heart disease is highly important as an example of early disease detection and aid in clinical decision-making. This paper proposes an effective machine learning technique for predicting heart disease, leveraging hyperparameter optimization and adaptive cross-validation. In this research paper, we considered an aggregated heart disease dataset comprising five well-known datasets from the UCI repository, totaling 1,190 records and 11 attributes. The eight classifiers applied in our study are: K-nearest neighbors, Naïve Bayes, Linear discriminant analysis, Quadratic discriminant analysis, Random forests, Support vector machine, Artificial neural network, and Decision tree. Hyperparameters were optimized through grid search and random search, and adaptive cross-validation with futility analysis was employed. From the results of the experiment, the optimized The K-Nearest Neighbors classifier performed better than all other classifiers with an accuracy of 98.32%, a precision of 97.66%, a recall of 99.21%, and an F1 score of 98.43%. The proposed The model demonstrates relatively good performance compared to other research using similar heart disease datasets and input attributes.

ABSTRACT. Sentiment analysis in the Arabic language faces numerous challenges due to the richness of the morphology in Arabic, the dialects, and the informal writing in social media sites, along with the lack of specialized data sets. In this paper, a comparison between classical machine learning methods and deep learning approaches will be conducted on a binary sentiment classification for Arabic YouTube comments. A manually labeled data set containing about 10,000 Arabic comments has been compiled and categorized as positive and negative sentiments. Five machine learning approaches have been chosen, namely Naive Bayes, Logistic Regression, Support Vector Machine, Multi-Layer Perceptron, and Random Forest, and three deep learning techniques, which are Simple Recurrent Neural Network, Convolutional Neural Network, and Long Short-Term Memory. An ensemble classifier model was also created based on majority voting to integrate the outputs from the individual classifiers. The designed ensemble model performed with a good in-domain accuracy score of 98.53%, along with improved cross domain generalization capabilities with a test accuracy of 63% on an external Arabic sentiment corpus. The trained model was also used as a web application by employing Streamlit technology to perform single-comment or batch sentiment analysis. It can be seen that ensemble learning is capable of producing a solid performance for Arabic sentiment classification, yet cross-domain remains a challenge for future work.

ABSTRACT. Abstract— The growing adoption of edge computing and Internet of Things (IoT) systems increases network traffic complexity, introducing major challenges in intrusion detection, such as privacy concerns, data heterogeneity (non-IID), and computational constraints. To address these issues, we propose Federated Learning Graph Convolutional Network Bidirectional Long Short-Term Memory (FedGCN-BiLSTM), a resilient FL framework. The architecture leverages GCNs with explicit adjacency matrices for spatial feature extraction, combined with Bidirectional LSTMs for temporal modelling. To mitigate client drift across heterogeneous networks (20 clients, three performance tiers), we employ FedProx regularization and FedNova step-normalized aggregation. The framework was evaluated on a comprehensive, merged dataset of 2.5 million samples from four public benchmarks (UNSW-NB15, IoT2025, EdgeIIoT, and Extra). Centralized evaluation demonstrates that the hybrid GCN-BiLSTM achieves 99.12% accuracy, 94.56% macro F1, and 0.986 MCC, outperforming CNN, LSTM, and sample-level GCN baselines. In the federated environment, it maintains a global accuracy of 99.08% and a binary anomaly detection AUC of 99.88%. To evaluate resilience against adversarial environments, we injected three synthetic attack scenarios: advanced malware, IoT botnets, and zero-day exploits. A unified Isolation Forest detector augmented with behavioural features (frequency, burst, and volatility) achieves a 99.9% detection rate. Following detection, a three-round adaptive recovery mechanism utilizing checkpoint restoration and fine-tuning with aggressive minority-class weights restores the system. This yields a statistically significant (p < 0.0001) improvement in the Matthews Correlation Coefficient (MCC) from approximately. 0.79 to 0.90 and raising the weighted F1-score above 0.94. All metrics, reported as mean ± std across 10 random seeds, show narrow confidence intervals that confirm structural robustness. Ultimately, FedGCN-BiLSTM provides a communication-efficient, highly accurate, and adaptive intrusion detection solution for heterogeneous edge environments.