SimST: A GNN-Free Spatio-Temporal Learning Framework for Traffic Forecasting

Xu Liu, Yuxuan Liang, Chao Huang, Hengchang Hu, Yushi Cao, Bryan Hooi, Roger Zimmermann

Keywords: Traffic Forecasting, Spatio-Temporal Graph Neural Networks

Abstract: Traffic forecasting is a crucial and challenging problem in smart city efforts. Spatio-Temporal Graph Neural Networks (STGNNs) have demonstrated great promise and become the de facto solution in this field. While successful, they require the message passing scheme of GNNs to construct spatial dependencies between nodes, and thus inevitably inherit the notorious inefficiency of GNNs. Given these facts, in this paper, we propose a simple yet effective GNN-free spatio-temporal learning framework, entitled SimST. Specifically, our framework replaces GNNs with two feasible and efficient spatial context injectors, which provide proximity and position information, respectively. SimST is also compatible with various temporal encoding backbones and involves a tailored training strategy. We conduct extensive experiments on five popular traffic benchmarks to assess the capability of SimST in terms of effectiveness and efficiency. Experimental results show that such a simple baseline performs surprisingly well. Using much fewer parameters, SimST not only achieves comparable or better performance than more sophisticated state-of-the-art STGNNs, but also obtains substantial throughput improvements.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)