Bayesian Knowledge Tracing (Two-State Hidden Markov Model)

Theory · All theories
Evidence · 14 claims (5 for, 6 mixed, 3 against) · 8 studies (3 causal, 2 design, 1 review, 1 associational, 1 theoretical), q1–q3 · 0 of 8 report an effect size · 14 claims rest on one study

Description

The survey describes Bayesian Knowledge Tracing (BKT), introduced by Corbett and Anderson, as a special case of the Hidden Markov Model with learning parameters (transition, forgetting) and performance parameters (guess, slip) plus an initial mastery probability. "BKT operates within a two-state student modeling framework: knowledge is either learned or unlearned, and there is no forgetting once a student has mastered the knowledge." Dynamic Bayesian knowledge tracing extends it to model hierarchies and relationships among knowledge concepts jointly.

Design Implications

Context

Requirements

• Per-knowledge-concept parameters: initial probability of mastery, transition probability, guessing probability and slipping probability
• Observed correct/incorrect answers used to update the posterior probability of mastery
• Response data tagged with the skills the instructor wants students to learn
• A history of student responses to problems (typically first attempts only)
• Values for the four probabilities L0, T, G, and S

Constraints

• Standard BKT models the parameters of each knowledge concept independently, although the survey notes knowledge concepts are hierarchical and closely related
• Fundamental models such as BKT often overlook forgetting
• The knowledge state is absorbent: the model assumes the student will not forget the skill once learned
• Usually a separate BKT model is fit for each skill

Target Learners

• Students whose exercise-answering interactions are recorded by online learning systems such as ASSISTments, Junyi Academy, Eedi and EdNet's Santa tutor
• students in online/cognitive-tutor courses (e.g., MOOC students)

Target Learning Objectives

• Monitoring students' evolving knowledge states on knowledge concepts and predicting performance on future exercises
• inferring mastery of instructor-tagged skills
• predicting future student performance

Claims

• Bkt Overestimates Accuracy After A Day Elapses [-W]
• Individualized Bkt Reduces Questions Needed For Mastery [+W]
• The survey reports, citing Desmarais and Baker, that students using the BKT-sequence recommendation algorithm solved more difficult exercises, obtained higher performance and spent more time in the system than students using the traditional approach. [+W]
• No credible spacing or color cue effects on the BKT learning rate were observed in the perceptual-cue reanalysis [~W]
• In simulated online mastery learning, BKT gave the fewest problems, Streak the second fewest, Random the most, and DKT gave almost as many as Random in one problem type and the least in the other two [~W]
• Adaptive scaffolding policies (BKT and DRL) significantly improve posttest performance over a non-adaptive control in a logic ITS [+W]
• Individualizing the initial knowledge parameter reliably improves prediction of real-world data [+W] — attached 2026-10-14 from Pardos et al. (2010), which proposed "Individualized Bayesian knowledge tracing framework".
• The exponential functional form of the BKT Markov chain calls into question the popular practice of fitting that model form to student data [-W]
• BKT students completed posttest problems significantly faster than Control students, with DRL showing marginal advantages in time and solution optimality [+W]
• Corbett and Anderson's original individualization attempts produced mixed results [~W]
• A Hidden Markov Model over knowledge states identifies Analytical Thinking (A5) as the learning bottleneck with the lowest forward transition probability [~W]
• Individualized Knowledge Tracing parameters fit student data significantly better and change mastery-learning practice decisions for many students [~W]
• Standard knowledge tracing models in use do not allow individual learning rates or individualized initial knowledge [-W]

The evidence by question

• Does choosing practice or scaffolding by BKT's mastery estimate improve learning or reduce the practice needed, compared with fixed or random sequencing, and for whom? Adaptive scaffolding policies (BKT and DRL) significantly improve posttest performance over a non-adaptive control in a logic ITS [+M]
• How accurately does BKT predict a learner's next answers, compared with other knowledge-tracing models, and do its assumptions and fitted parameters hold up? The survey states that deep learning knowledge tracing models exhibit superior performance to Bayesian and logistic models but have significant room for improvement in interpretability. [~M]

Related Theories

• Knowledge Tracing Model Taxonomy
• Mastery Learning
• Bayesian Knowledge Tracing: a four-parameter student-learning model in two forms (HMM and Knowledge Tracing Algorithm)
• Two hypotheses for the meaning of Spectral BKT's intermediate states: a stages-of-mastery account and a mastery-by-metacognition interaction account
• BKT+IRT: Bayesian Knowledge Tracing augmented with multidimensional generalizable student abilities and problem effects
• Bayesian Knowledge Tracing as a model of changing skill mastery during game-based assessment
• Intervention-BKT: a BKT extension that models the effect of instructional interventions on student knowledge states
• BKT Brute Force (BKT-BF): grid-search parameter estimation minimizing Residual Sum of Squares
• Spectral BKT: a BKT variant combining feature compensation (3-gram observations) with model compensation (four latent states)
• BKTransformer: transformer-based generation of temporally-evolving BKT parameters
• OptimNN: neural-network parameter generation (hypernetwork) for optimizing BKT

Examples

• Fit the combined parameter A and fix P(G) or P(L0) externally as an alternative to Dirichlet priors when fitting the BKT HMM

Key Sources

• Shuanghong Shen, Qi Liu, Zhenya Huang, Yonghe Zheng, Minghao Yin, Minjuan Wang, and Enhong Chen. (2021). A Survey of Knowledge Tracing: Models, Variants, and Applications. https://arxiv.org/abs/2105.15106
• Martori, F., Cuadros, J., & González-Sabaté, L. (2015). Direct estimation of the minimum RSS value for training Bayesian Knowledge Tracing parameters. Proceedings of the 8th International Conference on Educational Data Mining. https://www.educationaldatamining.org/EDM2015/proceedings/short364-367.pdf

Page metadata

Field	Value
type	theory
title	Bayesian Knowledge Tracing (Two-State Hidden Markov Model)
description	The survey describes Bayesian Knowledge Tracing (BKT), introduced by Corbett and Anderson, as a special case of the Hidden Markov Model with learning parameters (transition, forgetting) and performance parameters (gue...
status	review
canonical	True
generated	by claude/unspecified on 2026-09-24
trust	unverified — no one has confirmed this page's content
sources	2 — shuanghong-shen-2021, martori-2015

Frontmatter is the page's machine-readable record — see the Schema & Guide.