MCLKT: educational-objective-aware heterogeneous graph learning for multi-cognitive-level knowledge tracing
Year: 2026
Type: article
Abstract: Knowledge tracing (KT) commonly represents each knowledge concept by a latent mastery value, which does not directly distinguish the cognitive demand at which mastery is demonstrated. This study proposes Multi-Cognitive-Level Knowledge Tracing (MCLKT), a graph-based framework that represents each student-concept state using four ordered categories: not mastered, knowing, understanding, and applying. MCLKT separates exercise-concept incidence from exercise cognitive requirements through a binary Q mask and a graded P matrix, constructs a Student-Exercise-Knowledge (SEK) heterogeneous interaction graph, and uses a cumulative-link ordinal state head whose predicted states are directly coupled with the requirements of the next exercise for response prediction. Experiments were conducted on three datasets: CCNUC with 50 students, 124 exercises, and 17 knowledge concepts; T19 with 5,003 students, 207 exercises, and 11 knowledge concepts; and MR_sub with 3,950 students, 213 exercises, and 242 knowledge concepts. Under matched KK information, MCLKT-full outperformed GKT on AUC, ACC, and RMSE across all three datasets, while the no-KK variant remained competitive on CCNUC and most T19 measures. Full-P consistently outperformed both Q-only and Shuffled-P, indicating that graded cognitive requirements and their exercise-concept assignments provide additional predictive information beyond binary incidence. Independent CCNUC student-state annotations further provided ordinal agreement evidence, with the ordinal head achieving a QWK of 0.4505 ± 0.0496, while SPA and State × Requirement analyses provided complementary behavioral evidence. Overall, the results support the predictive value of heterogeneous interaction modeling and the validity of the learned ordered cognitive states. The source code and data are available at https://github.com/CCNUZFW/MCLKT .
Authors: Zhifeng Wang, Yulin Hou, Xiaoxue Liu, Chunyan Zeng
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: gold
Estimated APC paid: $1,990
Awards: G5243396600, G6698740542
OpenAlex ID: W7236727644
