Temporal Smoothness Doubly Robust Learning for Debiased Knowledge Tracing

Zhan, P., Chen, W., Chen, W., Pan, S., Cai, R. (2026) — arXiv preprint.

📄 Full text (arXiv)

Analysis

This paper addresses a critical but under-examined issue in KT systems: selection bias from non-random exercise recommendations. Prior KT methods train on observed logs using standard empirical risk, producing biased mastery estimates that compound errors in downstream recommendation loops. The proposed Temporal Smoothness Doubly Robust (TSDR) framework combines a propensity model with an error imputation model, regularized for temporal smoothness.^2605.05958

The doubly robust property ensures the estimator remains unbiased if either the propensity or imputation model is correct. This connects to personalized-learning by ensuring adaptive recommendations are not systematically biased toward high-engagement students.

Related Pages

• knowledge-tracing-irt — Doubly robust KT framework correcting selection bias

• learning-analytics — Temporal smoothness regularization for debiased student modeling

• personalized-learning — Unbiased mastery estimates for reliable adaptive recommendations

• intelligent-tutoring — Bias-corrected KT for exercise recommendation systems

• student-experience — Fair recommendation loops free from selection bias artifacts

Citation

APA: Zhan et al. (2026). Temporal Smoothness Doubly Robust Learning for Debiased Knowledge Tracing. arXiv:2605.05958. arXiv preprint.