Debiased Knowledge Tracing via Doubly Robust Learning
A new arXiv preprint (2605.05958) introduces a doubly robust learning method for Knowledge Tracing (KT) in intelligent education systems. KT relies on educational logs that suffer from selection bias due to non-random exercise recommendations and student choices. Existing methods ignore this bias, leading to inaccurate mastery estimates. The proposed approach integrates a propensity model with an error imputation model, ensuring unbiasedness if either model is correct. It also addresses variance-dependent stochastic deviations that accumulate over time, improving training stability and performance.
Key facts
- arXiv paper 2605.05958 proposes doubly robust learning for debiased Knowledge Tracing.
- KT in intelligent education systems relies on selectively observed educational logs.
- Non-random exercise recommendations and student choices cause severe selection bias.
- Existing KT methods neglect selection bias, yielding biased mastery estimates.
- The doubly robust formulation integrates a propensity model and an error imputation model.
- The method guarantees unbiasedness if either the propensity or imputation model is accurate.
- Variance-dependent stochastic deviations over time compromise estimator performance.
- A generalization bound is derived to mitigate training instability.
Entities
Institutions
- arXiv