BKTransformer: transformer-based generation of temporally-evolving BKT parameters¶

Theory · All theories
Evidence · no claims cited

Description¶

BKTransformer extends OptimNN by generating BKT parameters that vary over time within a student's problem sequence. A base parameter network f generates per-skill parameters, and a transformer decoder g (GPT-2-style, resembling SAKT) additively adjusts them conditioned on the student's sequence; the key architectural difference from SAKT is that the final layer "outputs BKT parameters as opposed to the correctness prediction". Predicting BKT parameters rather than correctness directly lets correctness predictions be "probabilistically justified by the produced slip, guess, forget, and learn parameters", retaining a layer of interpretability over deep KT methods while allowing non-Markovian learning and parameter sharing across skills.

Design Implications¶

Context¶

Requirements¶

• Penalty terms on adjustment magnitude and consecutive-timestep adjustment differences (coefficients lambda5, lambda6) to keep parameters from varying arbitrarily over time

Constraints¶

• The paper states it is unable to fully interpret the components of BKTransformer that precede the BKT layer, preventing understanding of how and why parameters evolve as they do

Target Learners¶

• students in intelligent tutoring systems and computerized tutoring contexts

Target Learning Objectives¶

• student response correctness prediction and mastery probability estimation in knowledge tracing

Claims¶

Related Theories¶

• Adaptive G-UKT: a unified probabilistic knowledge tracing framework coupling Gaussian state representations, adaptive graph topology learning, Wasserstein attention, and uncertainty-gated contrastive regularization
• Bayesian Knowledge Tracing: a two-state Hidden Markov Model inferring skill mastery from response histories
• Skill discovery BKT: end-to-end stochastic learning of the problem-KC assignment matrix with Gumbel-Softmax

Examples¶

-

Key Sources¶

• Badrinath, A. and Pardos, Z. (2023). Optimizing Bayesian Knowledge Tracing with Neural Network Parameter Generation. https://github.com/abadrinath947/OptimNN