StanBKT brings Bayesian inference to student knowledge tracing

Bayesian Knowledge Tracing (BKT) is a widely used student modeling approach in intelligent tutoring systems, but most implementations rely on point estimates from expectation-maximization, limiting uncertainty quantification. StanBKT, developed by Siddhartha Pradhan and five co-authors, rethinks this by providing an open-source Python package that leverages the Stan probabilistic programming language for full Bayesian inference. It supports Hamiltonian Monte Carlo, variational inference, Pathfinder, and optimization-based estimation, all while preserving the hidden Markov structure and interpretability of classical BKT. The package handles standard, grouped, and hierarchical models with flexible prior specification, posterior predictive checks, and diagnostic visualizations.

Evaluated on the large-scale ASSISTments 2020 dataset, StanBKT showed that different inference methods achieve comparable predictive performance but vary in computational efficiency and posterior fidelity. The researchers demonstrated how posterior inference enables principled comparisons of condition-specific parameters—learning, forgetting, guessing, and slipping—in an educational intervention involving perceptual cue manipulations. This uncertainty quantification facilitates more reliable interpretation of differences across experimental conditions, extending BKT beyond point estimation for robust probabilistic student modeling in educational data mining.