.png){kind=icon}
.png){kind=icon}
Return
Improving exercise-level Knowledge Tracing via Knowledge Concept-based Memory Network
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.127825.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge Tracing (KT), which works for learners to analyze learning states and predict future performance, is a critical task in intelligent tutoring systems. However, two significant issues have not been thoroughly investigated in KT research. First, most KT models achieve prediction at the knowledge concept (KC) level, as the sparsity of exercise data could lead to poor prediction performance. Although this issue has been addressed and exercise-level prediction is achieved via several KT models, the exercise embeddings acquired by these models fail to represent exercises reasonably. Thus, they cannot provide effective support for subsequent personalized recommendations. Second, another key problem is that current KT models mainly focus on improving prediction performance, neglecting the ability to trace knowledge states on all KCs. To tackle these two issues, we first introduce an exercise-level KT framework called the Knowledge Structure-aware Graph-Attention Network (KSGAN). In KSGAN, a graph attention layer and a representation optimization method are proposed to jointly train exercise representation vectors, achieving significant improvements in predicting learners' performance. Then, to evaluate the state of each KC, we extend KSGAN to a novel Knowledge Concept-based Memory Network (KCMN). This model quantifies the knowledge growth after each learning interaction, thereby capturing the change in mastery level for each KC. Extensive experiments on four datasets reveal that our models outperform some advanced KT models in prediction, as well as the superior ability of KCMN in evaluating the learner's knowledge states. The code can be found at https://github.com/syunnmo/KSGAN and https://github.com/syunnmo/KCMN.
Keywords:
Knowledge tracing
Graph attention networks
Memory networks
Personalized recommendation
Intelligent tutoring system
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W