A Two-Tower Spatial-Temporal Graph Neural Network for Traffic Speed Prediction
- Publisher:
- Springer International Publishing
- Publication Type:
- Chapter
- Citation:
- Advances in Knowledge Discovery and Data Mining, 2022, 13280 LNAI, pp. 406-418
- Issue Date:
- 2022-01-01
Closed Access
| Filename | Description | Size | |||
|---|---|---|---|---|---|
| 978-3-031-05933-9_32.pdf | Accepted version | 2.37 MB |
Copyright Clearance Process
- Recently Added
- In Progress
- Closed Access
This item is closed access and not available.
Recently, the remarkable effect of applying Dynamic Graph Neural Networks (DGNNs) to traffic speed prediction has received wide attention. Existing DGNN-based researches usually use a pre-defined or an adaptive matrix to capture the spatial correlations in traffic data. However, these static matrices are not enough to match the dynamic characteristics of spatial correlations. We argue that the global changes and local fluctuations of spatial correlations are dynamic with different frequencies. To this end, in this paper, we propose a Two-Tower DGNN (T2 -GNN) framework which divides the traffic data into a seasonal static component and an acyclic dynamic component, thus enhancing traffic speed prediction. The two components generated by an auto-decomposing block reflect global changes and local fluctuations of spatial correlations, respectively. Moreover, we use two parallel dynamic graph generation layers to construct a seasonal graph and an acyclic graph at each time step. In this way, the high-level representations of these two kinds of dynamic changes are learned through two dynamic graph convolution layers. Besides, the impact of fixed road network structure is modeled on the pre-defined graph and added to the spatial correlations. And we capture temporal correlations in temporal block before modeling spatial correlations. Finally, skip connections are used to converge the spatial-temporal correlations for final prediction. Experimental results on an urban dataset and two highway datasets show our proposed framework achieves the state-of-the-art prediction performances in terms of Mean Average Error (MAE) and Root Mean Squared Error (RMSE).
Please use this identifier to cite or link to this item: