Comparative Statistical Validation of Ensemble and Deep Learning Models for Enhanced Student Performance Prediction in Computer-Aided Learning Environments
TL;DR
Empirical findings conclusively demonstrate that ensemble and deep learning methods significantly surpass traditional statistical approaches in predictive accuracy and robustness in Computer-Aided Learning systems, and provide strong empirical support for integrating advanced ML models into CAL environments.
Abstract
This research conducts a rigorous statistical analysis of Computer-Aided Learning (CAL) systems, employing a diverse suite of machine learning (ML) algorithms to predict student academic performance. The study utilized a comprehensive dataset of 2,547 student records from academic years 2021–2024, integrating advanced preprocessing techniques such as SMOTE for class balancing and robust feature engineering. A comparative evaluation of seven distinct algorithms, including Logistic Regression, Support Vector Machines, Decision Trees, Random Forest, XGBoost, K-Nearest Neighbors, and Artificial Neural Networks (ANN), was performed using 10-fold stratified cross validation. Empirical findings conclusively demonstrate that ensemble and deep learning methods significantly surpass traditional statistical approaches in predictive accuracy and robustness. The Artificial Neural Network (ANN) emerged as the superior model, achieving the highest overall performance with an Accuracy of 90.1% and an ROC AUC of 0.950. Explainable AI techniques, specifically SHAP analysis, confirmed Prior-GPA (importance = 0.342), Assignment-Score (importance = 0.287), and Quiz-Score (importance = 0.246) as the most influential predictors of learning outcomes. These statistically validated insights provide strong empirical support for integrating advanced ML models into CAL environments to facilitate personalized learning, enable proactive early intervention strategies, and optimize educational resource allocation.