Temporal Graph Memory Networks For Knowledge Tracing

TL;DR

This work introduces Temporal Graph Memory Networks (TGMN) to address knowledge tracing by learning a unified embedding that jointly captures temporal exercise sequences and KC–KC relationships. It combines a temporal graph key-value memory with a dynamic graph convolution, a sequence-context GRU, and a memory-update mechanism including a forgetting-aware decay, using self-supervised pretraining for Question-KC embeddings. Empirical results on six KT benchmarks show state-of-the-art AUC performance with substantial gains over both sequence-only and graph-only baselines, supported by ablations and embedding analyses. The approach offers a scalable, generalizable framework for modeling forgetting and relational structure in student knowledge states, with potential impact on personalized tutoring systems.

Abstract

Tracing a student's knowledge growth given the past exercise answering is a vital objective in automatic tutoring systems to customize the learning experience. Yet, achieving this objective is a non-trivial task as it involves modeling the knowledge state across multiple knowledge components (KCs) while considering their temporal and relational dynamics during the learning process. Knowledge tracing methods have tackled this task by either modeling KCs' temporal dynamics using recurrent models or relational dynamics across KCs and questions using graph models. Albeit, there is a lack of methods that could learn joint embedding between relational and temporal dynamics of the task. Moreover, many methods that count for the impact of a student's forgetting behavior during the learning process use hand-crafted features, limiting their generalization on different scenarios. In this paper, we propose a novel method that jointly models the relational and temporal dynamics of the knowledge state using a deep temporal graph memory network. In addition, we propose a generic technique for representing a student's forgetting behavior using temporal decay constraints on the graph memory module. We demonstrate the effectiveness of our proposed method using multiple knowledge tracing benchmarks while comparing it to state-of-the-art methods.

Figures (6)

• Figure 1: An example for the temporal and structural dynamics involved in the Knowledge Tracing (KT) problem.
• Figure 2: Architecture of the temporal graph neural network (TGMN) model for knowledge tracing. Knowledge state estimation is performed by fusing long-term knowledge context $r_t$ (across all practice history) by addressing the temporal graph memory and the short-term knowledge context $h_t$ distilled from the current exercise sequence using a GRU cell. Afterward, the estimated knowledge state $m_t$ is fed into an answer prediction head to predict the correct answer probability for question $q_t$. We update the temporal graph memory (highlighted red arrows) guided by the status of the prediction error.
• Figure 3: Graph-based pretraining tasks used to learn useful questions and KCs embeddings. The left figure presents the objective of predicting the number of KC hops between a pair of question nodes. The right figure shows the objective of predicting the number of question hops given a pair of KCs.
• Figure 4: The Area under ROC curve (AUC) metric results comparing TGMN with the state-of-the-art KT models over all the datasets. We depict the average with the standard deviation of the metric values using a 5-fold splitting strategy (best in colors).
• Figure 5: The Area under ROC curve (AUC) training curves comparing four TGMN variants, each using a different embedding initialization.