Recovering Stranded Discrimination in Knowledge Tracing: Per-Item Bias Correction via Empirical-Bayes Shrinkage

In three linesSLC (State-space Logit Correction) corrects systematic per-item bias in deployed knowledge-tracing models. Using Laplace/IRLS transformation, empirical-Bayes shrinkage, and Kalman smoother, the method improves AUC across 4 datasets and 5 backbones, especially on sparse items. Global calibrators (Platt, temperature scaling) fail to recover lost discriminative ability.

Summary generated by Claude — human-verified