Predicting student performance using Moodle data and machine learning with feature importance

Abstract

Despite the growing technological advancement in education, poor academic performance of students remains challenging for educational institutions worldwide. The study aimed to predict students’ academic performance through modular object-oriented dynamic learning environment (Moodle) data and tree-based machine learning algorithms with feature importance. While previous studies aimed at increasing model performance, this study trained a model with multiple data sets and generic features for improved generalizability. Through a comparative analysis of random forest (RF), XGBoost, and C5.0 decision tree (DT) algorithms, the trained RF model emerged as the best model, achieving a good ROC-AUC score of 0.77 and 0.73 in training and testing sets, respectively. The feature importance aspect of the study identified the submission actions as the most crucial predictor of student performance while the delete actions as the least. The Moodle data used in the study was limited to 2-degree programs from the University of Southeastern Philippines (USeP). The 22 courses still resulted in a small sample size of 1,007. Future research should broaden its focus to increase generalizability. Overall, the findings highlight the potential of machine learning techniques to inform intervention strategies and enhance student support mechanisms in online education settings, contributing to the intersection of data science and education literature.