PAKT: New AI model improves knowledge tracing by 1.33% AUC
Phase-Aware Knowledge Tracing decomposes student behavior into ability and proficiency phases...
Knowledge tracing (KT) is the task of predicting students' future performance by modeling their evolving knowledge states from historical interactions. Traditional KT models treat the entire interaction sequence as a single behavioral process, overlooking phase-specific learning behaviors. Researchers from multiple institutions propose PAKT (Phase-Aware Knowledge Tracing), a framework that recognizes students transition from ability-building to proficiency-oriented learning after sufficient practice on previously failed concepts.
PAKT introduces a tailored decomposition mechanism to separate interactions into ability and proficiency phases, then uses a multi-branch Transformer with a type-aware readout module to jointly model phase-specific and holistic knowledge states. The paper also provides causal analysis revealing confounding bias in phase-agnostic models. On six public benchmarks, PAKT consistently outperforms all baselines, achieving a maximum AUC gain of 1.33% and an average gain of 0.82%. The work is available on arXiv as arXiv:2607.13103.
- Decomposes student interactions into ability and proficiency phases based on observed learning behavior transitions
- Uses a multi-branch Transformer with a type-aware readout module for joint phase-specific and holistic modeling
- Achieves up to 1.33% AUC gain over baselines, with 0.82% average improvement across six public benchmarks
Why It Matters
Better student performance prediction enables personalized learning pathways and earlier intervention in educational platforms.