Comparative Analysis of Machine Learning Models for Pedagogical Decision Support: Balancing Accuracy and Interpretability in Learning Analytics

Abstract

This article presents a comprehensive overview of a comparative analysis of three machine learning models  Decision Tree, Random Forest, and XGBoost  as tools for pedagogical decision support. As education becomes increasingly digitized, the focus of analytics is shifting from retrospective problem diagnosis to the proactive generation of actionable insights that help educators optimize the learning process. This study evaluates the models not only on their predictive accuracy but also on the interpretability of their results, a key factor for practical application in learning analytics. The methodology involved training and testing the models on a synthetic dataset simulating student behavior. The results demonstrated high predictive efficacy across all models, with accuracy ranging from 98% to 100%. The Decision Tree achieved 99% accuracy with complete logical transparency, making it a valuable tool for generating global pedagogical heuristics. XGBoost reached 100% accuracy, but its "black-box" nature necessitates the use of Explainable AI (XAI) methods for generating personalized insights. A key finding across all models was the high predictive power of behavioral indicators, such as the number of missed classes and assignment completion rates, confirming their importance in identifying students in need of support. This highlights the potential for such models to move beyond simple prediction and provide educators with targeted, evidence-based insights for timely intervention. The principal conclusion is the need for hybrid systems that synergistically combine the accuracy of complex models with the interpretability required for effective decision support in education.