GIKT: A Graph-based Interaction Model for Knowledge Tracing

Yang Yang, Jian Shen, Yanru Qu, Yunfei Liu, Kerong Wang, Yaoming Zhu, Weinan Zhang, Yong Yu

Abstract

With the rapid development in online education, knowledge tracing (KT) has become a fundamental problem which traces students' knowledge status and predicts their performance on new questions. Questions are often numerous in online education systems, and are always associated with much fewer skills. However, the previous literature fails to involve question information together with high-order question-skill correlations, which is mostly limited by data sparsity and multi-skill problems. From the model perspective, previous models can hardly capture the long-term dependency of student exercise history, and cannot model the interactions between student-questions, and student-skills in a consistent way. In this paper, we propose a Graph-based Interaction model for Knowledge Tracing (GIKT) to tackle the above probems. More specifically, GIKT utilizes graph convolutional network (GCN) to substantially incorporate question-skill correlations via embedding propagation. Besides, considering that relevant questions are usually scattered throughout the exercise history, and that question and skill are just different instantiations of knowledge, GIKT generalizes the degree of students' master of the question to the interactions between the student's current state, the student's history related exercises, the target question, and related skills. Experiments on three datasets demonstrate that GIKT achieves the new state-of-the-art performance, with at least 1% absolute AUC improvement.

Introduction

Knowledge Tracing (KT) is a critical task in personalized online education, aimed at modeling a student’s evolving mastery of concepts to predict future performance. Despite the success of deep learning in this domain, existing models often struggle with the dual challenges of question sparsity and the disconnect between specific questions and broader skills. Traditional models typically focus on broad skills (e.g., "Addition"), ignoring the nuanced difficulty of individual questions. Conversely, models focusing on specific questions suffer from data sparsity, as students only interact with a small fraction of the available question pool.

The GIKT Framework

To address these limitations, researchers from Shanghai Jiao Tong University and UIUC developed the Graph-based Interaction model for Knowledge Tracing (GIKT). GIKT bridges the gap between questions and skills by modeling their relationships as a bipartite graph and utilizing Graph Convolutional Networks (GCNs) for information propagation.

The overall architecture of the GIKT model, featuring RNN-based temporal modeling, GCN-based embedding propagation, and a history recap module.

Embedding Propagation via GCN

The core of GIKT is the representation of question-skill correlations. By treating questions and skills as nodes in a bipartite graph, the model uses GCN layers to propagate embeddings. This allows a question’s representation to be enriched by its required skills, and vice versa. This propagation mechanism effectively mitigates the question sparsity problem by allowing the model to learn from related skills even when question-specific data is limited.

Illustration of the bipartite graph structure between questions (q) and skills (s), showing how specific mathematical expressions map to broader concepts.

History Recap and Interaction Modules

Beyond graph-based embeddings, GIKT introduces a "History Recap" module. Unlike standard LSTMs that rely solely on the final hidden state, the recap module selectively retrieves relevant past exercises from the student's history. This prevents the model from "forgetting" distant but relevant interactions. Finally, a generalized interaction module performs pairwise comparisons between the student’s current state, retrieved history, the target question, and its associated skills to produce the final performance prediction.

Experimental Evaluation

The model was evaluated on three benchmark datasets: ASSIST09, ASSIST12, and EdNet. GIKT consistently outperformed state-of-the-art baselines, including DKT and DKVMN, achieving an absolute AUC improvement of at least 1% across all datasets.

Key Insights

  1. High-Order Correlations: Increasing GCN layers (up to 3) improved performance, indicating that "friend-of-a-friend" relationships between questions and skills are highly informative.
  2. Exercise-Level Retrieval: Retrieving specific past exercises proved more effective than retrieving past hidden states, as raw exercise data provides a cleaner signal for the current task.
  3. Consistency: Treating questions and skills as different instantiations of the same knowledge allows for a more robust and consistent interaction model.

Conclusion and Future Work

GIKT represents a significant advancement in KT by effectively integrating question-level information through graph-based methods. While the model shows superior performance, future research could explore incorporating additional side information, such as question text or temporal intervals between exercises, to further refine the student's knowledge state.

Summary updated

Abstract

With the rapid development in online education, knowledge tracing (KT) has become a fundamental problem which traces students' knowledge status and predicts their performance on new questions. Questions are often numerous in online education systems, and are always associated with much fewer skills. However, the previous literature fails to involve question information together with high-order question-skill correlations, which is mostly limited by data sparsity and multi-skill problems. From the model perspective, previous models can hardly capture the long-term dependency of student exercise history, and cannot model the interactions between student-questions, and student-skills in a consistent way. In this paper, we propose a Graph-based Interaction model for Knowledge Tracing (GIKT) to tackle the above probems. More specifically, GIKT utilizes graph convolutional network (GCN) to substantially incorporate question-skill correlations via embedding propagation. Besides, considering that relevant questions are usually scattered throughout the exercise history, and that question and skill are just different instantiations of knowledge, GIKT generalizes the degree of students' master of the question to the interactions between the student's current state, the student's history related exercises, the target question, and related skills. Experiments on three datasets demonstrate that GIKT achieves the new state-of-the-art performance, with at least 1% absolute AUC improvement.
BibTeX
@inproceedings{yang2020gikt,
  author = {Yang Yang and Jian Shen and Yanru Qu and Yunfei Liu and Kerong Wang and Yaoming Zhu and Weinan Zhang and Yong Yu},
  title = {{GIKT: A Graph-based Interaction Model for Knowledge Tracing}},
  booktitle = {ECML/PKDD},
  year = {2020},
  doi = {10.1007/978-3-030-67658-2\_18},
}
Figure