An Explainable Graph Enhanced Ensemble for Confusion Detection and Multiple Strategy Intervention in Massive Open Online Course Discussion Forums

Abstract

—Learner confusion in Massive Open Online Course (MOOC) forums is a critical precursor to dropout. Existing de tection methods often treat posts in isolation, ignoring structural discussion dynamics. We present ConFusionGraph, an end-to-end framework integrating confusion detection, explainable analysis, and personalized intervention. The methodology models forum interactions as heterogeneous graphs with five relation types capturing thread structure, authorship, and temporal dynamics. We employ a stacked ensemble combining gradient boosting, transformer-based language models, and a heterogeneous graph neural network. To ensure transparency, dual local and global feature attribution analyses reveal that while explicit flags domi nate, structural features—such as thread mean confusion and re ply degree—are top contributors. Furthermore, attribution-based clustering identifies two primary mechanisms: active questioning and confusion contagion. Experiments on a large-scale dataset (nearly 30,000 posts across 11 courses) demonstrate an accuracy of 89.2% and an AUC of 0.923. Confusion chain analysis provides empirical evidence of propagation through threads and cross thread contagion. These findings drive a multi-strategy interven tion system offering peer answer retrieval, thread summarization, and personalized clarification. This provides a robust, context aware mechanism to mitigate learner dropout through timely support.

Keywords