New Method Improves Knowledge Tracing Accuracy

Deployed knowledge-tracing models often suffer from systematic per-item logit bias, which arises from limited expressivity in their core architectures and shifts in item properties post-deployment. This bias significantly degrades prediction quality. While global post-hoc calibrators like Platt scaling can improve probability estimates, they do not enhance the model's discriminative ability, as measured by AUC, because they are monotone score-only transformations. To address this, a new method called State-space Logit Correction (SLC) has been introduced. SLC aims to recover "stranded discrimination" by explicitly conditioning on item identity. The process involves converting binary observations into Gaussian pseudo-observations using Laplace/IRLS, applying empirical-Bayes shrinkage via a Kalman smoother, and then fitting an offset-Platt link. The state-space formulation also provides a detectability bound, explaining why temporal tracking offers no additional benefit at current data densities. Across four datasets, five backbone architectures, and multiple seeds, SLC consistently improved AUC on all datasets and NLL on three, with its most significant advantages observed for sparse items. This phenomenon, where entity-level bias is left unaddressed by the backbone, is suggested to extend beyond educational contexts.