Open Access

Article

Article ID: 4181

Predictive model for students’ academic performance using machine learning approach

by Wadzani Aduwamai Gadzama, Ogah Stephen Ugbowu, Lucy Bulus Dalhatu

Computing and Artificial Intelligence, Vol.4, No.1, 2026;

The early prediction of students' academic performance using machine learning has emerged as a valuable approach for identifying at-risk learners and enabling timely intervention. Many factors, such as students' academic background, prior performance, institutional policies, and learning environment, influence educational outcomes; their complex interplay remains inadequately understood in many contexts. This study aimed to explore the effectiveness of machine learning algorithms in predicting students' academic performance at Adamawa State University, Mubi, Adamawa State, Nigeria. This study used 1,730 datasets from the academic records of first-year students from the Faculty of Science for the 2024/2025 academic session. The study split the datasets into 80% training and 20% for testing. Data were analysed using the Waikato Environment for Knowledge Analysis (Weka) and Python. The model was evaluated using students' cumulative grade point averages (CGPAs) from the academic session results. The machine learning algorithms used were Logistic Regression (LR), Random Forest (RF), Decision Trees (DT), Naïve Bayesian (NB), and Support Vector Machines (SVM). Experimental results based on various performance metrics indicate that the SVM model achieved the best result with an accuracy of 0.92, precision of 0.92, recall of 0.93, and F1-score of 0.93. The results revealed that the SVM approach outperforms individual benchmark methods and provides robust insight into factors that determine academic success. The findings offer evidence-based guidance for educators, departments, faculties, institutional management, and policymakers to design targeted interventions to improve learning outcomes.

Open Access

Article

Article ID: 4104

Performance evaluation of B-Tree and hash indexing under varying data sizes in relational database systems

by Hassan Bediar Hashim

Computing and Artificial Intelligence, Vol.4, No.1, 2026;

This study investigates query performance optimization in relational database management systems (RDBMSs) by evaluating two common indexing techniques, B-Tree and Hash indexing, under varying dataset sizes. With the rapid growth of data generated by IoT systems, enterprise applications, and digital services, efficient query execution has become essential for maintaining scalability and system performance. The research compares three database configurations: no indexing, B-Tree indexing, and Hash indexing, while applying a Cost-Based Optimization (CBO) strategy to improve query plan selection. Experimental results reveal that query response time increases significantly with larger datasets, especially when no indexing is used. Both indexing methods substantially enhance performance compared to full-table scans, achieving improvements ranging from 35% to 60% depending on dataset size and query workload. The measured speedup factors reached up to 2.60×, confirming the effectiveness of indexing in reducing execution time. Further analysis indicates that B-Tree indexing consistently performs better than Hash indexing in large-scale and mixed-query environments due to its logarithmic search efficiency and support for range queries. B-Tree indexing reduced execution time to nearly 40–45% of the baseline, whereas Hash indexing achieved approximately 55–60% under similar conditions. The findings emphasize that selecting an appropriate indexing strategy is critical for optimizing database query performance, and that the effectiveness of each method depends largely on workload characteristics and dataset scale.

Open Access

Review

Article ID: 3991

Image processing techniques for detection of objects in blurry pictures: A comprehensive review

by Ghassan Abdullah Abdulwasea Al-Maamari, Mubarak Mohammed Al-Ezzi Sufyan, Ramzi Hamid Abdo Al-Jaberi, Mokhtar H. Al-Sarori, Mahfoudh Al-Asaly, Asma’a Khalil Alkershi

Computing and Artificial Intelligence, Vol.4, No.1, 2026;

Detecting objects in blurry and degraded images remains a critical unsolved challenge in computer vision, affecting applications from medical diagnostics and autonomous navigation to remote sensing and surveillance. Image degradation caused by motion, defocus, poor lighting, or environmental factors severely compromises feature visibility and limits the performance of conventional detection algorithms. This paper presents a comprehensive, systematic review of state-of-the-art techniques designed to address this problem. We first categorize common image degradations and analyze classical and deep learning-based solutions for image deblurring and enhancement, including CNN (Convolutional Neural Networks), GAN (Generative Adversarial Networks), and transformer architectures. The review then critically examines object detection models, particularly YOLO (You Only Look Once) and CNN-based networks, adapted for low-quality inputs. A key focus is on integrated pipelines that jointly optimize restoration and detection. We synthesize findings from over 200 studies, highlighting performance across diverse domains such as UAV (Unmanned Aerial Vehicle) imagery, underwater exploration, and medical analysis. Furthermore, we discuss standard datasets and evaluation metrics, identify persistent challenges including real-time processing, multi-degradation handling, and domain adaptation, and outline promising research directions. This review serves as a foundational resource for researchers and practitioners aiming to build robust vision systems for real-world, blur-prone environments.

Open Access

Article

Article ID: 3958

From calculators to artificial intelligence: A multi-level framework for technology adoption and resistance in education and organisations

by Michael Mncedisi Willie

Computing and Artificial Intelligence, Vol.4, No.1, 2026;

Technological innovations are frequently associated with enhanced efficiency and improved decision-making; however, their initial adoption has often been characterised by notable resistance. This study examined the persistence of this phenomenon by synthesising empirical evidence from the historical adoption of calculators, spreadsheets, and statistical software, and by comparing these insights with contemporary developments in artificial intelligence (AI). The findings indicated that resistance extended beyond technical limitations and reflected a complex, multi-dimensional process shaped by individual factors such as self-efficacy, perceived usefulness, and anxieties related to skill displacement as well as organisational culture and broader systemic conditions. The evidence further demonstrated that successful adoption was contingent upon structured support mechanisms, incremental exposure to new technologies, and the clear articulation of value propositions. At the organisational level, effective leadership, adequate resource allocation, and coherent policy alignment emerged as critical enablers of sustained integration. Drawing on these insights, the study proposed a multi-level conceptual framework that integrates individual, organisational, and systemic determinants to guide future technology adoption. The framework underscores that meaningful and sustainable integration depends on coordinated, cross-level interventions rather than isolated initiatives. The framework also highlights the need for further empirical research into AI adoption in under-resourced contexts. It calls for critical engagement with the ethical implications of algorithmic decision-making in evolving socio-technical systems.

Open Access

Article

Article ID: 4261

Next-generation VANET routing: An AI-based time-evolving graph framework

by Ali Hamlili

Computing and Artificial Intelligence, Vol.3, No.4, 2025;

Vehicular Ad Hoc Networks (VANETs) exhibit highly dynamic topologies due to rapid vehicle mobility and frequent fluctuations in wireless connectivity. Consequently, efficient routing remains a significant challenge. Proactive routing protocols rely on periodic route updates. This creates a trade-off between routing accuracy and control overhead. Higher update frequencies improve accuracy and timeliness. However, they also generate substantial control overhead. Conversely, lower update frequencies reduce overhead at the cost of outdated topology information and increased routing errors. Achieving an optimal balance between these competing objectives is particularly difficult in non-stationary network environments, a well-documented challenge in VANETs. This paper introduces a unified framework combining stochastic modeling, graph analysis, and machine-learning optimization. Vehicle positions follow a Poisson process, while link dynamics are modeled with a two-state Markov chain. We derive closed‑form expressions that quantify network connectivity, expected link lifetime, and overall routing stability in dynamic vehicular environments. An artificial intelligence (AI)-driven adaptive controller is integrated into the proposed framework to dynamically adjust routing parameters in response to variations in VANET topology and non-stationary network conditions. Simulation results demonstrate a 35–50% reduction in routing overhead while preserving high packet delivery performance.

Open Access

Article

Article ID: 4397

Leveraging extensive feature modeling for facial emotion recognition

by Milica Tufegdzic, Nevena Tufegdzic, Marija Mojsilovic

Computing and Artificial Intelligence, Vol.3, No.4, 2025;

Facial emotion recognition (FER) is an important area of affective computing with applications in human–computer interaction, healthcare, education, and intelligent systems. Although recent FER research is largely dominated by deep learning and transformer-based approaches, handcrafted feature modeling remains attractive due to its interpretability and lower computational requirements. This study proposes an Action Unit–based machine learning (AU-ML) framework for recognizing basic emotions from lateral facial expressions using the Karolinska Directed Emotional Faces (KDEF) dataset. Facial Action Units (AUs) were extracted, and manual feature selection was performed to retain only AU intensity and presence information relevant to emotion recognition. This process significantly reduced the original feature vector, improving computational efficiency while preserving classification performance. To compensate for the reduced dataset size after extracting lateral images, data augmentation techniques, including horizontal flipping, shifting, scaling, rotation, and brightness and contrast adjustments, were applied prior to AU extraction. Several machine learning algorithms were evaluated, including K-Nearest Neighbors, Support Vector Classifier, Decision Tree, Naïve Bayes, Random Forest, AdaBoost, Bagging, Voting, and Stacking Classifiers, CatBoost, and Extreme Gradient Boosting. The results demonstrated that ensemble methods generally outperformed simpler classifiers, with CatBoost achieving the best classification performance. The findings indicate that extensive feature modeling remains a reliable approach to emotion recognition, and that AU-based representations provide an interpretable and computationally efficient alternative to deep learning approaches.