Asian Journal of Research in Computer Science

Bayesian Knowledge Tracing Driven Adaptive Learning for Personalized GCE Examination Preparation

Blessed Nkurumah Kpudzeka — 2026-08-28

Examination preparation for the General Certificate of Education in Cameroon relies heavily on static past question booklets that treat every learner identically regardless of what that learner has already mastered. This paper reports the design, implementation, and simulation-based validation of a Bayesian Knowledge Tracing-driven adaptive quiz engine built for GCE Ordinary Level Chemistry within the ToriLearn prototype platform. The engine estimates a learner's mastery of eleven syllabus-defined knowledge components after every response, using the canonical four-parameter Bayesian Knowledge Tracing formulation together with penalty adjustments for hint use, skipped questions, and viewed answers, and it feeds the resulting mastery estimate into a real-time decision engine that selects the next question's difficulty. We describe the mathematical formulation as implemented, the sequence of operations executed on every answer submission, and three synthetic learner scenarios constructed to test whether the engine's behaviour matches the behaviour the underlying model predicts. Results show that a consistently correct simulated learner reaches high mastery within a small number of attempts and is advanced to harder material, that a consistently incorrect learner is retained on remedial content without mastery collapsing to zero, and that a learner with alternating performance produces a fluctuating difficulty trajectory that tracks the fluctuating mastery estimate. The endpoint that performs this computation responded within 45 to 80 milliseconds under local development testing. No real student data were used to fit the model parameters and no classroom trial was conducted, so the contribution of this paper is a validated implementation of the Bayesian mechanism rather than measured evidence of examination performance improvement.

Intelligent Models for Mitigating the Impact of Phishing on Electronic Commerce Data: A Critical Narrative Review of Detection Capability, Robustness and Operational Evidence

Chukwueke Nwagbara — 2026-08-26

Phishing remains the dominant entry point through which the personal, financial and transactional data held by electronic commerce platforms are compromised, and the research response has been overwhelmingly computational. Several hundred studies now propose intelligent models based on feature-engineered classifiers, deep neural architectures, multimodal reference-based systems and, most recently, large language models. Reported classification performance is consistently high, yet losses attributable to online shopping fraud and credential theft have not declined correspondingly. This review examines that discrepancy. Its purpose is to evaluate critically what the available evidence establishes about the ability of intelligent models to reduce data loss in electronic commerce settings, as distinct from their ability to separate labelled samples within curated datasets. Literature was identified through open scholarly databases and citation searching, appraised for methodological adequacy, and synthesised thematically around five problems: the mapping between detection outputs and the data assets actually at risk; the comparative evidence for competing model families; the construction and temporal validity of evaluation datasets; robustness under adversarial and distributional pressure; and the conditions under which a detection decision becomes a mitigation. The evidence indicates that headline performance figures are strongly conditioned by dataset construction, that accuracy on balanced benchmarks translates poorly to the extreme base-rate asymmetry of live traffic, and that robustness has been assessed for only a small proportion of published models. Reference-based and multimodal designs show more stable behaviour under impersonation than purely lexical models, but at a computational cost that is rarely reported in terms compatible with transaction-time constraints. Language-model detectors improve semantic sensitivity and explanation quality while introducing latency, cost and privacy trade-offs that remain sparsely characterised. Evidence linking model deployment to reduced data compromise is almost entirely absent. Priorities include temporally partitioned and platform-realistic evaluation, standardised adversarial reporting, and outcome measures defined in terms of data exposure rather than classification alone.

Dual Text-guided Cross-attention for Global–local Visual Fusion in Vietnamese Visual Question Answering

Huy Tran — 2026-08-25

Visual Question Answering (VQA) requires effective cross-modal reasoning between visual content and natural-language questions. This challenge is particularly significant for Vietnamese due to the relatively limited availability of annotated VQA resources. This study proposes a dual-stream architecture that exploits complementary Transformer-based and convolutional visual representations for Vietnamese VQA. A Vision Transformer (ViT) is employed to obtain globally contextualised visual features, while ConvNeXt V2 preserves spatially structured visual information. Questions are encoded using PhoBERT. Instead of directly merging the two visual streams, the proposed model uses the textual representation as a shared query in two independent cross-attention modules, allowing question-relevant information to be retrieved separately from each visual representation prior to fusion. The resulting representations are concatenated along the sequence dimension, pooled, and combined with the sentence-level textual embedding through a residual connection for answer classification. Experiments on the ViVQA benchmark show that the proposed architecture achieves 64.31% accuracy and 62.65% F1-score. Additional experiments across multiple ConvNeXt V2 backbone scales consistently show improvements over the corresponding single-stream convolutional baselines, supporting the complementary role of the two visual representations.

Implementation of Partial Cross Entropy Loss for Point-supervised Remote Sensing Image Segmentation

Loraine Mutune — 2026-08-25

Weakly supervised semantic segmentation has emerged as a promising approach for reducing the annotation burden associated with dense pixel-level labelling. Among weak supervision strategies, point supervision offers an attractive compromise by requiring only a small subset of labelled pixels while preserving meaningful semantic information. However, effective optimisation under sparse supervision requires loss functions capable of excluding unlabelled regions from the training process. This study presents an implementation framework for point-supervised remote sensing image segmentation using Partial Cross Entropy Loss. Dense segmentation masks obtained from the LoveDA dataset were converted into sparse point annotations through random pixel sampling with a point ratio of 1%. A subset containing 200 image-mask pairs was constructed and partitioned into training and validation sets. A U-Net architecture with a ResNet-34 encoder was employed as the segmentation backbone, while Partial Cross Entropy Loss was implemented using the ignore-index mechanism available in PyTorch to restrict optimization to labelled pixels only. Functional validation was performed through forward propagation, loss computation, and gradient backpropagation. Successful parameter updates and finite loss values confirmed the correct integration of sparse point supervision with the encoder-decoder segmentation network. The proposed framework provides a simple, modular, and reproducible implementation for point-supervised semantic segmentation and establishes a foundation for future investigations involving weakly supervised and semi-supervised remote sensing image analysis.

AI-Based Traffic Violation Detection System Using Computer Vision

Oluwasefunmi Busola Famodimu — 2026-08-27

Aims: This project presents the development and implementation of an AI-based traffic violation detection system designed to automate the monitoring of road traffic violations using CCTV footage and computer vision techniques. The aim is to create a comprehensive solution that can analyse real-time or recorded video streams to detect violations such as red-light running and speeding, thereby improving road safety and reducing the manual surveillance burden.

Study Design: The research employed a modular, data-driven approach combining software engineering best practices with advanced machine learning and computer vision techniques. The system was developed iteratively with a component-based architecture supporting separation of concerns and scalability.

Place and Duration of Study: The study was conducted at Babcock University, Department of Computer Science, School of Computing, Ilishan-Remo, Nigeria, from September 2023 to May 2024.

Methodology: The implementation utilised Python as the primary programming language, OpenCV for image processing and video analysis, and the YOLOv8 deep learning model for real-time object detection and vehicle classification. Multi-object tracking was accomplished using the DeepSORT algorithm, enabling consistent vehicle identification across video frames. Speed estimation was performed through frame displacement analysis and timestamp calculations. A rule-based violation detection module was integrated to identify traffic offences, including red-light violations and speeding. An Automatic Number Plate Recognition (ANPR) component utilising Optical Character Recognition (OCR) was included for vehicle identification. A structured MySQL database was implemented to store violation records with timestamps and evidence. An administrative dashboard, developed using HTML, CSS, JavaScript, and Bootstrap, provides real-time analytics, processing history, and violation reporting capabilities. The system architecture follows a three-tier model: Client-Side (Frontend), Server-Side (Backend), and Database Layer.

Results: Functional testing confirmed successful video upload, accurate vehicle detection using the YOLOv8 model, which demonstrated high precision, reliable multi-object tracking with consistent vehicle ID maintenance, accurate speed estimation through frame analysis, and correct violation flagging with real-time visual alerts. Performance evaluation demonstrated that the YOLOv8 model processed video frames efficiently at optimal detection speed, maintaining consistent detection accuracy and tracking reliability under various traffic conditions, including different lighting, weather, and traffic density scenarios. The system successfully generated violation reports with vehicle identification through licence plate recognition, timestamps, and evidence imagery for enforcement purposes.

Conclusion: The AI-based traffic violation detection system provides a practical and innovative solution to contemporary traffic violation control and management challenges. By integrating advanced computer vision features, including object detection, multi-object tracking, and automatic number plate recognition, the system analyses captured footage with high accuracy to identify violations and extract relevant vehicle details. The system's ability to provide real-time evidence capture and generate automated violation reports demonstrates how modern AI technologies can be effectively leveraged to create practical tools that meet contemporary traffic management needs, enabling authorities to manage road safety with greater effectiveness and efficiency.

Neural Network-based Model for Detecting Credit Card Fraud: A Comparative Study of Oversampling Techniques and Feature Selection

Kalala Kanyinda Norbert — 2026-08-29

Credit card fraud detection is complicated by the severe class imbalance typical of transaction data, because fraudulent cases represent only a small proportion of observations. This study develops a neural-network-based model for classifying transactions as legitimate or fraudulent and compares combinations of two oversampling techniques and two feature-selection approaches. The dataset contains 20,000 observations and 26 variables, with 339 fraudulent transactions (1.7%). After removing the transaction identifier and transforming categorical variables, class imbalance was addressed using the Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic Sampling (ADASYN). Relevant features were then selected using Pearson’s correlation coefficient or XGBoost, producing four configurations: SMOTE + Pearson, SMOTE + XGBoost, ADASYN + Pearson, and ADASYN + XGBoost. The models were evaluated using accuracy, precision, recall, specificity, F1 score, loss, mean squared error, training and validation curves, and receiver operating characteristic analysis. Although the ADASYN + Pearson configuration produced high nominal accuracy and recall, its zero specificity and lower F1 score indicated poor identification of legitimate transactions. In contrast, the ADASYN + XGBoost configuration showed a more balanced performance across the evaluation criteria, with accuracy, precision, recall, specificity, and F1 score each reported at 79%. These results support ADASYN + XGBoost as the best-performing configuration among the four models evaluated in this study.

Development of a Web-based Academic Advising Platform Leveraging Extreme Gradient Boost (XGBoost) Algorithm

Kazeem O. N. — 2026-09-01

Many tertiary institutions continue to rely on fragmented, manual, and reactive academic-advising processes, which may delay access to timely student support. This study developed and evaluated a web-based academic advising platform integrating Extreme Gradient Boosting (XGBoost) with student-facing academic-support functions. The platform combines a student dashboard, GPA calculator and simulation tool, graduation-eligibility checker, course catalogue, resource hub, results-management functions, and an automated AI-supported advising interface. Academic records were obtained from Al-Hikmah University, the University of Ibadan, and the Open University Learning Analytics Dataset repository, yielding a reported total of 5,550 records. The data were preprocessed, partitioned into training, validation, and hold-out test sets, and used to train and evaluate the XGBoost model for student-performance classification and academic-risk identification. The reported evaluation produced 92% accuracy, 90% precision, 94% recall, 91% specificity, and an F1-score of 92%. The platform was designed to support early identification of students who may require academic assistance while improving access to academic information and decision-support functions. The findings indicate that integrating machine learning with a web-based advising environment can support data-informed academic monitoring and advising. Further multi-institutional validation and real-world deployment evaluation are required to determine the system’s generalisability and operational performance across different institutional contexts.

A Low-Cost Multimodal Pothole Detection Framework Using YOLO-Based Vision and Infrared Depth Sensing for Road Infrastructure Maintenance

Yasheena Niromi Kaumini — 2026-09-03

Aims: Road potholes are a leading cause of vehicle damage, reduced road safety, and increased accident risk, particularly in developing countries where road maintenance resources are limited. Existing automated detection systems rely predominantly on single-modality approaches, either vision-based or sensor-based, that cannot simultaneously localise potholes spatially and estimate their physical depth. This study presents the first low-cost, edge-deployed multimodal pothole detection framework that combines a deep learning visual detector with an analogue infrared sensor array for real-time depth estimation on commodity embedded hardware.

Methodology: A YOLOv12m object detection model was trained on 4,211 road images and paired with a 12-channel Sharp GP2Y0A21 infrared sensor array. Both subsystems were deployed on a Raspberry Pi 5 platform connected via two Arduino UNO microcontrollers. A decision-level weighted confidence fusion algorithm combined their outputs. An SVM classifier trained on 32 engineered IR features provided a four-level severity classification (None, Minor, Moderate, Severe). The SVM was evaluated using a corrected 5-fold stratified cross-validation pipeline to ensure leakage-free results.

Results: YOLOv12m achieved Precision = 1.0000(at the operational confidence threshold), Recall = 0.9952, and mAP@50 = 0.9950 at an inference latency of 4.13 ms per frame, outperforming six comparative YOLO architectures on all primary metrics. The SVM severity classifier achieved cross-validated accuracy = 0.9900, F1-score = 0.9831, and ROC-AUC = 1.000 across five folds. The fused system operates at 25 Hz on the Raspberry Pi 5 platform.

Conclusion: To the best of our knowledge, no prior published work has combined a spatially distributed multi-channel analogue infrared sensor array with a real-time YOLO visual detector on a single low-cost edge platform for simultaneous pothole localisation and physical severity estimation. The framework provides actionable, severity-graded outputs suitable for proactive road maintenance planning and directly contributes to reducing pothole-related road safety hazards.