Assessing Predictive Model Portability Through Algorithmic Approaches in Educational Data Mining

Abstract

This study investigates the portability of predictive models across Moodle-based courses using behavioral data collected during the Covid-19 lockdown from four university courses at Ibn Tofail University. It evaluates three types of machine learning approaches: distance-based models (K-Nearest Neighbors and Support Vector Machines), ensemble methods (Random Forest and XGBoost), and probabilistic models (Naive Bayes). Models are trained on individual courses and directly applied to others without retraining. Their performance is assessed using standard classification metrics alongside portability-specific indicators such as the Performance Degradation Rate (PDR) and the Portability Index (PI). Results demonstrate that ensemble methods consistently achieve the highest transferability across courses. In contrast, distance-based and probabilistic models tend to show lower generalization, potentially due to interaction sparsity and grade imbalance. These findings underscore the importance of algorithmic selection in designing scalable and transferable predictive systems in educational data mining.