Student Performance Prediction With Large-Scale xAPI and SIS Data: Feature Importance and Ensemble Comparison

Abstract

The early identification of students who may be at risk requires predictive models that can integrate behavioral and institutional data while maintaining reliable performance across academic terms. This study examines the relative predictive information carried by xAPI behavioral traces and Student Information System (SIS) academic context variables and compares three ensemble learning paradigms across four cumulative quarters of the term. The analysis used a large-scale institutional dataset comprising over 18 million xAPI interaction records, more than 111,000 student-course observations, and over 3,400 heterogeneous course instances across five terms. xAPI behavioral features and SIS context variables were combined into a unified feature set, and Bagging, Boosting, and Stacking models were evaluated at each temporal stage. Behavioral features showed the strongest individual feature-level signal, with assignment-related interactions emerging as the most informative predictors under the information-theoretic feature importance metrics. Among the ensemble approaches, XGBoost achieved the highest area under the ROC curve (AUC) at every temporal stage, increasing from 0.847 in the first quarter to 0.932 in the fourth. The descriptive AUC ordering remained stable across all four quarters, with XGBoost achieving the numerically highest AUC, followed by the evaluated Stacking configuration and Random Forest. All three ensemble paradigms exceeded an AUC of 0.82 after the first quarter of the term, suggesting that integrated xAPI and SIS data may support early risk screening before mid-term.