MS-BKT: a multistate knowledge tracing model with data-driven recency weights replacing the learning rate¶

Theory · All theories
Evidence · no claims cited

Description¶

MS-BKT is a knowledge tracing architecture that keeps BKT's HMM structure but makes two changes: the knowledge node expands from 2 to 21 states, and a recency weight parameter R replaces the transition probability p(T), so that "weight increases incrementally by a fixed quantum R for each new attempt" with the optimal value learned from data. Knowledge is updated per state via Bayes' rule and estimated by the MAP mode of the posterior, with parameters fit by Expectation Maximization. The authors position it as retaining BKT's parameter interpretability while balancing recent and historical data.

Design Implications¶

Context¶

Requirements¶

• Requires estimating the free recency parameter R and other parameters from data during model fitting
• The recency weight quantum R must be learned from data during model fitting

Constraints¶

• The choice of 21 states was made as granular enough with manageable calculation overhead; the authors note the number of states can be explored further, including a continuous distribution function
• Prediction quality on held-out data was similar to, not better than, classic BKT on most datasets

Target Learners¶

• K-12 students practicing tagged skills in intelligent tutoring systems
• Students in intelligent tutoring systems using mastery learning

Target Learning Objectives¶

• Real-time estimation of student mastery of a skill to drive adaptive learning
• Accurate, interpretable real-time estimation of student knowledge level

Claims¶

Related Theories¶

• Two hypotheses for the meaning of Spectral BKT's intermediate states: a stages-of-mastery account and a mastery-by-metacognition interaction account
• Bayesian Knowledge Tracing as a model of changing skill mastery during game-based assessment

Examples¶

-

Key Sources¶

• Agarwal, D., Baker, R.S., & Muraleedharan, A. (2020). Dynamic knowledge tracing through data driven recency weights. Proceedings of The 13th International Conference on Educational Data Mining (EDM 2020). https://educationaldatamining.org/edm2020/