.png){kind=icon}
.png){kind=icon}
Return
Dynamic heterogeneous graph-structured knowledge tracing with multi-behavior integration and temporal forgetting
.png){kind=icon}
OA
AI
DOI：10.1038/s41598-026-66887-2.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge tracing (KT) aims to estimate the temporal evolution of learners’ knowledge states and predict future exercise performance. Most existing KT methods focus on exercise-response sequences and therefore cannot explicitly distinguish knowledge exposure through learning resources from knowledge verification through problem solving. We propose Dynamic Heterogeneous Graph-Structured Knowledge Tracing (DHGKT), a framework that organizes learners, exercises, videos, concepts, and optional comments within a unified relational event representation. Each learner is associated with an interpretable concept-mastery vector. Before each event, an exponential temporal decay is applied; video events provide a duration-aware exposure signal, whereas exercise events calibrate the state through prediction-error feedback. Semantic concept-video weighting and concept-aligned social exposure are treated as auxiliary extensions rather than indispensable components. Under a chronological 70%/10%/20% training/validation/test split with five random seeds on four MOOCCubeX courses, DHGKT-Core achieved an average test area under the receiver operating characteristic curve (AUC) of 0.8529, compared with 0.8318 for its exercise-only variant and 0.7508 for a same-input multi-behavior fusion control. The contribution of video behavior was course dependent: it improved three courses but reduced performance on C_801420. Fixed and learnable topology-aggregation variants produced negligible AUC changes, and the semantic and social extensions did not yield reliable additional gains. These results indicate that the principal value of DHGKT lies in transparent event-driven knowledge-state evolution rather than generic graph convolution. The revision further clarifies the roles of behavioral modalities, graph-event semantics, decay sensitivity, and evaluation limitations in response to the second-round comments.
Keywords:
Knowledge tracing
Heterogeneous graph-structured modeling
Multi-behavior learning
Temporal forgetting
Interpretable knowledge state
Learning analytics
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
3.9
Papers:
27.1W
Citations:
83.5W