A REPRODUCIBLE EXPLAINABLE AI PIPELINE FOR TEACHER-FACING STUDENT DIGITAL TWINS

Abstract

Learning management systems record large volumes of student activity, yet their built-in analytics tend to describe the past rather than predict outcomes. We ask a deliberately falsifiable question: do engineered Student Digital Twin features improve grade prediction over a competent baseline drawn from ordinary LMS signals, and are the resulting explanations stable enough to put in front of a teacher? To answer it we built a leakage-aware pipeline that pairs a weekly student-state representation with a gradient-boosting predictor and a perturbation-based explanation layer, then ran a nested feature ablation on public data from two institutions. The result isn't encouraging. Added feature richness does not help in any reliable way: no Twin block beats the baseline by more than one RMSE point on the primary cohort, only one cell shows a clear gain, and explanation rankings shift with the evaluation regime (Kendall τ between 0.32 and 0.79).