DyGKT: Dynamic Graph Learning for Knowledge Tracing

TL;DR

DyGKT introduces a continuous-time dynamic graph approach to knowledge tracing, addressing infinite data growth, irregular time intervals, and evolving student-question-concept relations. By constructing a subgraph around each interaction, encoding historical neighbor sequences with a dual time encoder, and using a multiset indicator to capture evolving structural relationships, the model updates time-aware representations for students and questions via continuous-time GRUs. A link-level classifier then predicts performance with an optimized cross-entropy objective, and extensive experiments across five real-world datasets demonstrate superior AP and AUC against a broad set of baselines, with ablation confirming the contributions of MI and dtE. The work provides a scalable, inductive KT framework suitable for dynamic educational environments and offers practical insights into temporal and structural factors shaping student learning trajectories.

Abstract

Knowledge Tracing aims to assess student learning states by predicting their performance in answering questions. Different from the existing research which utilizes fixed-length learning sequence to obtain the student states and regards KT as a static problem, this work is motivated by three dynamical characteristics: 1) The scales of students answering records are constantly growing; 2) The semantics of time intervals between the records vary; 3) The relationships between students, questions and concepts are evolving. The three dynamical characteristics above contain the great potential to revolutionize the existing knowledge tracing methods. Along this line, we propose a Dynamic Graph-based Knowledge Tracing model, namely DyGKT. In particular, a continuous-time dynamic question-answering graph for knowledge tracing is constructed to deal with the infinitely growing answering behaviors, and it is worth mentioning that it is the first time dynamic graph learning technology is used in this field. Then, a dual time encoder is proposed to capture long-term and short-term semantics among the different time intervals. Finally, a multiset indicator is utilized to model the evolving relationships between students, questions, and concepts via the graph structural feature. Numerous experiments are conducted on five real-world datasets, and the results demonstrate the superiority of our model. All the used resources are publicly available at https://github.com/PengLinzhi/DyGKT.

Figures (11)

• Figure 1: Knowledge tracing with dynamic graph to trace students' learning states. This task has three dynamics: 1) constantly growing scale; 2) varying semantics of time intervals between long and short time intervals; 3) evolving relationships between students, questions, and concepts.
• Figure 2: We develop a continuous dynamic graph learning model for KT. (a) We first convert input into a continuous dynamic graph, then extract the historical first-hop interaction of both the student node and the question node. (b) Next, embed the interaction neighbor sequence with a performance encoder, dual time encoder, and multiset indicator. (c) Then, we stack the two sequences with their respective features and feed them separately into two sequential models. (d) Finally, the outputs of the sequential models are used for the knowledge tracing task.
• Figure 3: Bar chart depicting the distribution of statistical time intervals for five datasets. It can be found that there are a large number of students answering questions continuously in the real data, and then starting another question after some time between breaks, which once again confirms the logic of our dual time encoder module.
• Figure 4: Visualization of a student's knowledge mastery degree of one question over 15 steps in Assist17; the boxes with different colors in the time interval line represent different time interval ranges between the current and next steps; the darker-colored circles indicate that the attempted question belongs to the predicted link multiset.
• Figure 5: AP for transductive ablation study of DyGKT.

Theorems & Definitions (2)

• definition 1
• definition 2