.png){kind=icon}
.png){kind=icon}
Return
Multiple consistency constraints for knowledge tracing
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ipm.2026.105108.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Existing knowledge tracing (KT) models may achieve strong predictive performance by capturing non-knowledge-related statistical patterns in the data, without needing to faithfully model students’ knowledge states, thereby introducing the risk of biased knowledge state modeling. To address this issue, we propose a novel KT model, <strong>M</strong>ultiple <strong>C</strong>onsistency <strong>C</strong>onstraints for <strong>K</strong>nowledge <strong>T</strong>racing (MCCKT), designed to improve the accuracy and cognitive plausibility of knowledge state modeling while further enhancing predictive performance. MCCKT constructs factual–contrastive sample pairs via controlled behavioral perturbations and employs perturbation-based consistency regularization to impose the proposed multiple consistency constraints. These constraints ensure that the model adheres to the monotonicity assumption when modeling students’ answering states, including knowledge states, test-taking psychological states, and ideal performance, thereby encouraging the model to learn answering-state representations that conform to educational cognitive principles. In addition, MCCKT designs a hierarchical prediction mechanism from ideal performance to factual performance. Building on ideal performance jointly determined by knowledge states and test-taking psychological states, this mechanism explicitly accounts for the effects of students’ guessing and slipping, resulting in predictions that more closely reflect real-world answering performance. Experimental results on three educational datasets demonstrate that MCCKT achieves competitive performance in terms of both AUC and ACC compared with mainstream baseline models. Furthermore, we introduce a targeted evaluation metric, Group AUC of Mastery (GAUCM). The GAUCM results provide supplementary evidence that the knowledge state representations learned by MCCKT better conform to the monotonicity assumption, offering indirect support for their cognitive plausibility.
Keywords:
Knowledge tracing
Multiple consistency constraints
Monotonicity assumption
Test-taking psychological state
Consistency regularization
Journal
I
IF:
6.9
Papers:
310
Citations:
0