Spatiotemporal fusion knowledge tracking model based on spatiotemporal graph and fourier graph neural network
Abstract Knowledge Tracing (KT) aims to predict students’ future learning performance, which mainly involves dynamic changes in both temporal and spatial dimensions. The temporal dimension captures dynamic evolution of knowledge acquisition (e.g., accumulation/forgetting), and the spatial dimension...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-07-01
|
| Series: | Journal of King Saud University: Computer and Information Sciences |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44443-025-00138-8 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Knowledge Tracing (KT) aims to predict students’ future learning performance, which mainly involves dynamic changes in both temporal and spatial dimensions. The temporal dimension captures dynamic evolution of knowledge acquisition (e.g., accumulation/forgetting), and the spatial dimension models associations between knowledge points. Current state-of-the-art Graph Neural Network (GNN)-based methods typically require spatial networks (e.g., Graph Convolutional Network) to capture static spatial dependencies between knowledge points and temporal networks (e.g., Long Short-Time Memory) to model local temporal dependencies in the learning sequence. However, the uncertain compatibility of these two networks imposes an additional burden on model design, and the separated spatiotemporal modelling violates the spatiotemporal inter-dependencies of the real-life learning process, leading to the shortcomings of existing models in accurately predicting the state of learners’ knowledge. To solve the problem of uncertain spatiotemporal compatibility, this paper proposes a knowledge tracing model that unifies spatiotemporal information - Spatiotemporal Fourier Knowledge Tracing(STFKT). The model constructs spatiotemporal graphs by integrating information from the time dimension and the spatial dimension through information extracted from the cognitive and behavioral perspectives, resolves the compatibility problem between the two, and processes spatiotemporal dependencies features in the frequency domain through Fourier Graph Neural Network (FourierGNN) to capture complex spatiotemporal relationships, and improve computational efficiency and accurate modeling of spatiotemporal features. Experimental results show that STFKT outperforms KT models such as DKT, SAKT, and GKT across multiple datasets. In particular, it achieves an AUC improvement of 19.53%–38.68% on the ASSISTments2017 dataset, demonstrating notable predictive performance in scenarios with complex knowledge structures and long-term dependencies. |
|---|---|
| ISSN: | 1319-1578 2213-1248 |