Answers are grounded in this paper and cite their sources.
Knowledge tracing (KT) models are usually compared by ranking metrics, yet adaptive systems act on predicted probabilities, which may degrade when the response history is short or noisy. This study audits next-response probabilities of five predictors (a smoothed item-accuracy baseline, Bayesian knowledge tracing with forgetting, 2PL item response theory, deep knowledge tracing (DKT), and a simplified attention-based model), holding the predicted events fixed while varying visible history length (3–50 events), history label flips (0–20%), and guessing and slipping separately. Learners were simulated from two generators with different assumptions. We report ROC-AUC, Brier score, negative log-likelihood, and expected calibration error with learner-level bootstrap intervals, plus high-confidence errors, and compare temperature with Platt scaling. Model ordering depended on the generator and on item information: under the dynamic generator no model beat the item baseline on Brier score at three visible events, and there DKT with item outputs matched the attention-based model. Truncating a long record made every history-based model under-predict (by up to 0.086), because each began the window from a state fitted to earlier practice; when three events were a learner’s complete record, the same models were nearly unbiased. Temperature scaling, which has no intercept, left this offset in place in most cases and transferred inconsistently from long histories, whereas Platt scaling fitted in the target condition removed it. Label flips lowered discrimination and, with short histories, mostly erased the advantage over the baseline. The findings are a controlled reference for validation on real interaction logs.
No takes yet. Share an insight, caveat, or question.
Meixian Jin (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: