Interpretable Dual-Branch Framework for Dropout Prediction in E-Learning Using Static Profiles and Temporal Engagement

Abstract

Student dropout remains a persistent challenge in large-scale digital learning plat- forms, where early identification of at-risk learners is critical. We propose a dual-branch pre- diction framework that combines interpretable static features with sequential behavioral traces. A Random Forest branch captures demographic and aggregate engagement factors, while an LSTM branch models week-by-week activity dynamics. The fused representation balances ac- curacy and transparency, enabling both high predictive performance (AUC = 0.997) and clear attribution of risk drivers. Beyond accuracy, the study reveals that static engagement fea- tures remain the strongest predictors of dropout, while temporal patterns provide early-warning signals that enrich intervention timing. Unlike prior approaches that emphasize either inter- pretability or raw sequence modeling, our design integrates both perspectives into a unified framework, making dropout prediction both actionable and deployment-ready. Experiments on two benchmark datasets confirm the complementary roles of static and temporal signals, and demonstrate how the proposed model supports real-time, explainable interventions in digital learning environments.