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

The paper introduces SLC (State-space Logit Correction), a novel method for correcting per-item bias in knowledge tracing models that typically suffer from logit bias post-deployment. SLC employs a state-space formulation to convert binary observations into Gaussian pseudo-observations, utilizes empirical-Bayes shrinkage via a Kalman smoother, and fits an offset-Platt link, resulting in improved AUC across multiple datasets and architectures. This approach is significant for practitioners as it enhances model discriminative ability, particularly for sparse items, addressing limitations in existing calibration methods that do not rectify underlying biases.