Bayesian Deep Knowledge Tracing: A Hybrid Bayesian-Neural Approach to Student Modeling
Year: 2026
Type: conference-paper
Abstract: Adaptive tutoring systems depend on accurate, trustworthy estimates of student knowledge—yet most existing knowledge tracing models produce point predictions that carry no measure of their own reliability. When a system cannot distinguish a confident prediction from an uncertain one, it risks triggering interventions that are pedagogically inappropriate, particularly in early or data-scarce stages of a learning sequence. This work addresses that limitation by introducing Bayesian Deep Knowledge Tracing (BDKT), a hybrid framework in which latent knowledge states are represented as posterior distributions over a multi-skill space and propagated through time via an LSTM-based transition model. Unlike deterministic tracers such as DKT, and unlike attention- or graph-based extensions that improve predictive accuracy without addressing calibration, BDKT couples sequential neural modeling with Monte Carlo dropout to quantify both epistemic uncertainty—stemming from limited data—and aleatoric uncertainty inherent in student responses. A graph-informed multivariate Gaussian covariance further encodes prerequisite dependencies among skills, a structural dimension absent from prior neural KT models. Evaluated on a large-scale public dataset of over 500 000 interactions spanning 30 mathematical skills, BDKT yields statistically significant improvements over DKT in discrimination (AUC 0.87 vs. 0.83, p<0.01) and reduces expected calibration error by nearly half (ECE 0.04 vs. 0.07, p<0.01), supporting the practical value of uncertainty-aware student modeling.
Institutions: Université de Fianarantsoa, University of Antananarivo
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: hybrid
OpenAlex ID: W7201870751