Multimodal Predictive Modelling in Education: A Comparative Evaluation of ML, Boosting, DL, and Hybrid Fusion Architectures

Abstract

The rapid growth of online learning platforms has generated large volumes of educational data that can support student performance prediction. However, existing Machine Learning (ML) and Deep Learning (DL) approaches often face challenges related to generalizability, feature integration, and early identification of at-risk students. This study presents a comparative evaluation of classical ML, boosting-based, DL, and a proposed Hybrid Artificial Neural Network-Deep Fusion Model (ANN-DFM) using the Open University Learning Analytics Dataset (OULAD). A unified experimental framework incorporating data cleaning, feature selection, and class balancing was employed to ensure fair benchmarking. Results show that Logistic Regression and Support Vector Machines achieved moderate accuracy (0.68-0.76), while Random Forest reached 0.91 accuracy. Boosting models, including XGBoost, LightGBM, and CatBoost, improved performance to 0.94-0.95 accuracy. Among DL approaches, ANN achieved 0.93 accuracy. The proposed ANN-DFM outperformed all baseline models, achieving 96.65% accuracy and an F1-score of 0.97, while demonstrating stable early-quarter predictions. The findings highlight the effectiveness of multimodal feature fusion for enhancing predictive accuracy and supporting early educational interventions.