.png){kind=icon}
.png){kind=icon}
Return
Uncertainty-aware structural trend-decoupling knowledge tracing
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.116879.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge tracing (KT) infers students’ evolving mastery from interaction sequences and serves as a core component of intelligent tutoring systems. Although recent Transformer-based and probabilistic KT models achieve strong predictive performance, they often entangle long-term learning trends with short-term behavioral fluctuations, while inadequately modeling forgetting dynamics under irregular real-time intervals. To address these issues, we propose UST-KT, an uncertainty-aware probabilistic knowledge tracing framework. Specifically, UST-KT introduces a structural cognitive decoupling module that employs causal convolution to separate smooth cognitive trends from high-frequency fluctuations, injecting them into the mean and covariance of Gaussian latent states, respectively. In this way, the mean captures stable long-term mastery, whereas the covariance serves as a proxy for fluctuation-related uncertainty. Furthermore, we develop a Hawkes–Wasserstein attention mechanism. This mechanism modulates Wasserstein-distance-based distributional similarity using a Hawkes-inspired continuous-time decay kernel, enabling more time-aware modeling of forgetting dynamics under irregular time gaps. Extensive experiments on three real-world benchmark datasets demonstrate that UST-KT consistently outperforms strong baselines and yields more stable prediction dynamics, better-calibrated predictive probabilities, together with more structured covariance representations that serve as proxies for fluctuation-related unpredictability.
Keywords:
Knowledge tracing
Educational data mining
Uncertainty modeling
Temporal dynamics
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W