Disentangling Knowledge States with Ability and Proficiency Modeling for Knowledge Tracing

Traditional Knowledge Tracing (KT) models often treat student learning as a uniform process, failing to distinguish between the initial acquisition of a concept and the later consolidation of skills. This paper asks: can we improve performance prediction by explicitly modeling these two distinct phases of learning?

The authors introduce Phase-Aware Knowledge Tracing (PAKT), a framework that segments a student's interaction history into two phases: the ability phase and the proficiency phase.

• Ability Phase: Captures the initial period where a student is learning a new concept, primarily reflecting their cognitive aptitude.
• Proficiency Phase: Captures the later period after mastery, where performance is driven by practice, fluency, and skill automation.

PAKT uses a decomposition mechanism based on a cumulative correct response threshold to partition the data. These partitioned sequences are then processed by a multi-branch Transformer architecture. This backbone includes separate decoders for the ability, proficiency, and holistic (complete) sequences, which are then integrated by a type-aware readout module to generate final predictions.

PAKT was evaluated against several representative baselines across six public educational datasets. The results demonstrate that the phase-aware approach consistently outperforms standard, phase-agnostic models. Specifically, the proposed method achieved a maximum AUC gain of 1.33% and an average gain of 0.82% across the benchmarks. The authors also provide a causal analysis suggesting that the performance gains stem from reducing the confounding bias inherent in mixing early-stage acquisition signals with late-stage proficiency signals.

This research highlights that learning is not a monolithic process. By acknowledging that the factors influencing a student's performance change as they move from novice to expert, the model provides more accurate and interpretable predictions. This approach offers a practical way to enhance Intelligent Tutoring Systems (ITS) by tailoring feedback to the specific phase of a student's learning journey.