Artificial neural network modelling of student’s academic performance based on online learning activities pattern

Early and accurate prediction of student academic performance is critical for enabling timely interventions. This research proposes an intelligent classification framework that models academic performance using real-time engagement data collected from an online learning platform during the first eight (8) weeks out of 17 weeks of a programming course. The study addresses four core challenges in predictive academic modelling: reliance on outdated static data, class imbalance in small dataset, underutilization of multi-activity features, and suboptimal hyperparameter tuning in neural network architectures. A dataset of 99 samples with 12 academic engagement features was collected, pre-processed using Min-Max normalization, and balanced using the Adaptive Synthetic Sampling (ADASYN) algorithm with synthetic data was statistically verified using the Kolmogorov–Smirnov test, histogram comparisons, and boxplot analyses. Feature selection was conducted using Mutual Information, Recursive Feature Elimination (RFE), Random Forest Importance, L1 Regularization Least Absolute Shrinkage and Selection Operator (L1 Lasso), and eXtreme Gradient Boosting (XGBoost) to identify the optimal predictors. A Multilayer Perceptron (MLP) classifier was developed and optimized using five optimizers: Adaptive Moment Estimation with Decoupled Weight Decay (AdamW), Nesterov-accelerated Adaptive Moment Estimation (Nadam), Adaptive Moment Estimation with Maximum of Past Squared Gradients (AmsGrad), Adaptive Gradient Algorithm (AdaGrad), and Stochastic Gradient Descent with Momentum (SGD with Momentum). Experimental results demonstrate that RFE-selected features combined with adaptive optimizers significantly enhance model generalization and stability. The MLP consistently outperformed other architectures including Support Vector Machine (SVM), k-Nearest Neighbors (kNN), CNN, Recurrent Neural Network (RNN), and Long Short-Term Memory Network (LSTM), achieving an F1-score of 90.0% and a testing accuracy of 86.7%, affirming its robustness for non-sequential educational data. Key outcomes from this research include ADASYN-integrated MLP for improved minority-class sensitivity, feature-efficient framework (six selected features via RFE), and empirical optimizer comparison (AdamW, Nadam, AmsGrad, AdaGrad, SDG-Momentum) This research advances scalable academic early-warning systems aligned with Malaysia’s Outcome-Based Education (OBE) and Continuous Quality Improvement (CQI) initiatives, offering a real-time, pedagogically actionable tool for educators.