Disentangling Knowledge States with Ability and Proficiency Modeling for Knowledge Tracing

Short summary

This paper proposes Phase-Aware Knowledge Tracing (PAKT), which decomposes student interactions into ability-building and proficiency-oriented phases rather than treating all interactions uniformly. It uses a multi-branch Transformer with type-aware readout to capture phase-specific knowledge states and provides causal analysis of confounding bias in phase-agnostic models. Experiments on six benchmarks show consistent improvements with up to 1.33% AUC gain over baselines.

• •Introduces PAKT, decomposing student learning into ability and proficiency phases
• •Uses multi-branch Transformer with type-aware readout for phase-specific knowledge modeling
• •Outperforms baselines on six benchmarks with up to 1.33% AUC gain

Generated with AI, which can make mistakes.