Structured Inductive Bias for Multi-timescale Knowledge Tracing

Abstract

Knowledge tracing (KT) aims to infer learners' evolving cognitive states from interaction sequences to predict future performance. However, learning behaviors are inherently non-stationary and multi-timescale, where long-term cognitive accumulation and short-term contextual fluctuations are tightly entangled. Most existing KT models address this complexity by modeling different timescales as parallel feature channels, implicitly assuming their independence. This architectural bias overlooks the hierarchical and co-evolutionary nature of learning dynamics, often leading to degraded generalization under limited learning evidence. To address this limitation, we propose Frequency-aware Structured Knowledge Tracing (FSKT), which introduces Experiential Learning Theory (ELT) as a structured inductive bias to explicitly model the co-evolution of multi-timescale cognitive dynamics. Specifically, FSKT first applies a causal stationary wavelet transform to decompose latent representations into low-frequency components capturing long-term cognitive accumulation and high-frequency components reflecting short-term contextual fluctuations. Building upon this decomposition, we design a cascaded cognitive evolution architecture aligned with the four ELT stages (Experience? Reflection? Knowledge? Application), introducing an inductive ordering over information flow across frequencies and stages. To further disentangle stage-specific semantics, FSKT employs geometric projection and residual stripping mechanisms, enabling each stage to selectively extract and refine information from distinct frequency bases. Additionally, a weakly supervised sufficiency constraint is introduced to encourage alignment between stage representations and their corresponding observable learning signals, improving structural consistency. Extensive experiments on 5 datasets demonstrate that FSKT achieves competitive performance while exhibiting more stable generalization under data sparsity and varying observed history lengths. The code is available at https://github.com/Oia-10/FSKT.