Abstract:

Early warning systems in higher education frequently treat grades as direct indicators of academic ability, ignoring behavioral processes such as course withdrawal and disengagement that distort observed performance. We propose a dual-process framework that models academic outcomes as the interaction between latent capacity and behavioral truncation. The framework separates academic potential from administrative and behavioral effects, enabling structurally grounded prediction. To operationalize this perspective, we compare a Principal Component Analysis (PCA) baseline with an optimized Coupled Matrix Factorization (CMF) architecture designed to learn capacity independently from zero-inflation. A controlled simulation environment with known generative parameters allows evaluation against theoretical ground truth. Results show that CMF achieves substantially lower prediction error, preserves distributional structure, and maintains individualized trajectory patterns, whereas PCA produces homogenized estimates and fails to represent discontinuities introduced by withdrawal penalties. By distinguishing capacity deficits from behavioral instability, the framework supports differentiated intervention strategies and more equitable decision-making. These findings reposition early warning systems as pedagogically grounded decision-support infrastructures and highlight the importance of representational validity in AI-driven educational analytics.

Keywords:

Educational Data Generation, Student Performance Prediction, Coupled Matrix Factorization, Machine Learning.