.png){kind=icon}
.png){kind=icon}
Return
Multi-view spatio-temporal graph convolutional network with domain adversarial learning for traffic prediction
.png){kind=icon}
OA
AI
Z
W
DOI：10.1016/j.ijtst.2025.10.018.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning methods have been widely applied in urban traffic flow prediction and have achieved promising results. However, these methods rely on large amounts of training data. In reality, due to the scarcity of traffic data in some cities, deep learning methods struggle to achieve optimal predictive performance. Recently, transfer learning has been introduced into traffic flow prediction to alleviate data scarcity through knowledge transfer. Nevertheless, most existing studies primarily focus on modeling spatial adjacency within traffic networks, overlooking the inherently multidimensional spatial structures of urban transportation and the distribution discrepancies between cities. In this paper, we propose a multi-view spatio-temporal graph convolutional network (MSTGCN) with domain-adversarial learning to address the challenges of traffic flow prediction under data-scarce conditions. Specifically, MSTGCN constructs a multi-view spatio-temporal graph based on the physical distance, functional similarity, and dynamic correlation of traffic nodes. On this basis, a parameter-sharing spatio-temporal graph convolution is designed to extract spatio-temporal features of traffic flow. Additionally, a domain-adversarial learning approach is integrated to extract domain-invariant spatio-temporal features and reduce distribution differences. Extensive experiments on five real-world traffic datasets demonstrate that MSTGCN significantly outperforms state-of-the-art transfer learning methods, achieving up to 9.7% lower mean absolute error on average.
Keywords:
Traffic flow prediction
Transfer learning
Multi-viewspatio-temporal graph convolutional network (MSTGCN)
Domain adversarial learning
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
4.8
Papers:
1.5K
Citations:
1.5K
Organization
Cited Papers