Predicting Learner Disengagement in E-Learning Platforms Using Interpretable Machine Learning: A Comparative Study of Tree-Based Classifiers

Student disengagement is a significant obstacle that blocks effective results in E-learning during the digital education time. The research employs a machine learning methodology to accurately categorize student disengagement levels into Low, Medium, and High, utilizing data from 14,101 student records that incorporate 21 characteristics encompassing demographic information, student behavior, and course engagement data. Traditional classifiers became unusable because the data contained mixed numerical and categorical fields, yet required intensive data encoding. Our research focused on two highly effective tree-based models, Random Forest and XGBoost, because they automatically work with data containing various types. Testing demonstrated that XGBoost outperformed Random Forest, achieving a $94 \%$ accuracy rate that exceeded the $93 \%$ precision and recall in all disengagement stages. Random Forest delivered 83% accuracy, yet its recall performance was only $65 \%$ for medium disengagement cases. The Random Forest AUC $=0.95$ and XGBoost AUC $=0.99$. Therefore, the proposed model demonstrates outstanding discrimination performance, as its AUC value approaches 1.0, indicating superior distinction between disengagement categories with minimal instances of misclassification. XGBoost offers an accurate disengagement forecasting method that preserves data interpretability by avoiding information distortion induced by encoding. The study provides practical solutions that enable E-learning systems to identify struggling learners before deploying targeted intervention approaches to enhance student success rates across courses. Research teams will concentrate on developing explainable intervention methods that utilize temporal modeling strategies and perform cross-platform validation.