In deployed knowledge tracing models, the models are typically frozen post-training; however, systematic per-item logit bias arises due to limited expressivity in backbone architectures and shifts in item properties after deployment, degrading prediction quality. While global post-hoc calibrators such as Platt scaling, temperature scaling, and isotonic regression improve probability estimates, they do not affect discriminative ability as measured by AUC. This invariance of AUC is a structural consequence of monotone score-only transforms; recovering stranded discrimination requires conditioning on item identity.
We propose SLC (State-space Logit Correction), which converts binary observations to Gaussian pseudo-observations via Laplace/IRLS, applies empirical-Bayes shrinkage through a Kalman smoother, and fits an offset-Platt link. The state-space formulation also yields a detectability bound that characterizes the Bernoulli information floor, explaining why temporal tracking provides no benefit at current data densities. Across four datasets, five backbones, and three seeds, SLC improves AUC on all four datasets and NLL on three, with the advantage concentrating on sparse items. Cross-domain controls suggest that the same phenomenon can arise beyond education when the deployed backbone leaves entity-level bias.
Blogger's Review: The SLC method proposed in this paper significantly enhances the discriminative ability of knowledge tracing models through per-item bias correction, particularly in sparse datasets. By combining state-space theory with empirical-Bayes methods, it showcases innovative ideas and practical effects, offering insights for applications beyond the education domain.