Knowledge tracing via multiple-state diffusion representation

doi:10.1016/j.eswa.2024.124797

2024-12-01

PRE

DOI：10.1016/j.eswa.2024.124797

原文链接

原文求助

摘要

En 中文

Knowledge tracing aims to supervise students' grasping of concepts, inferring their knowledge states, and predicting their future performance. Current research predominantly models knowledge states to represent the overall concept mastery of students, reflecting their understanding of learning content. However, knowledge states exhibit diverse changes in different cognitive scenarios. This singular form of knowledge state is challenging to authentically reflect the changing process of concept states in real learning environments. To address this issue, we propose a knowledge-tracing model with a multiple-state diffusion representation. This model captures students' knowledge states in various cognitive modes, thereby representing the resting state and deducing the executing state in different tasks. Specifically, we first construct three relationship structures in exercises for initialization. Next, in the diffusion process guided by the classifier, we represent the resting state and extract relationships of concept states to capture the influence of historical exercises. Finally, deducing the executing state is done by incorporating channel attention. Through experiments comparing the proposed model with four relevant models on three datasets, the model has demonstrated certain advantages in terms of AUC, ACC, and representation quality, indicating its effectiveness. Furthermore, by incorporating the influence of historical exercises, the model describes the internal changes in concept states and their relationships, refining the evolutionary logic of knowledge states in different learning scenarios. This highlights the model's interpretability.

Keyword:

Knowledge tracing

Knowledge states

Diffusion model

Classifier guidance