Interpretable Student Performance Mining from E-Learning Interaction Logs Using Feature Selection, Classification, and Clustering

Abstract

Educational assessment generally evaluates students using final marks, while technology-enhanced learning environments can record the learning actions that precede those marks. This research presents an Educational Data Mining framework for analyzing student performance from interaction logs generated by MathTutor, an e-learning application for mensuration. Five classifiers - Naive Bayes, Multilayer Perceptron, ZeroR, J48, and Random Forest - were compared. J48 and Multilayer Perceptron achieved 100% reported accuracy; J48 required only 0.03 seconds to build the model compared with 27.19 seconds for the Multilayer Perceptron and was selected as the preferred interpretable classifier. Expectation Maximization clustering produced five groups whose membership counts matched the five grade categories generated by J48. The findings further indicate that high-performing learners generally viewed more examples, practiced more exercises, and used fewer hints. The study demonstrates how interpretable data mining can supplement score-based assessment with evidence about learning behaviour and provide a basis for targeted remedial support.