.png){kind=icon}
.png){kind=icon}
Return
Spatio-temporal graph mixformer for traffic forecasting
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2023.120281.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Traffic forecasting is of great importance for intelligent transportation systems (ITS). Because of the intricacy implied in traffic behavior and the non-Euclidean nature of traffic data, it is challenging to give an accurate traffic prediction. Despite that previous studies considered the relationship between different nodes, the majority have relied on a static representation and failed to capture the dynamic node interactions over time. Additionally, prior studies employed RNN-based models to capture the temporal dependency. While RNNs are a popular choice for forecasting problems, they tend to be memory hungry and slow to train. Furthermore, recent studies start utilizing similarity algorithms to better express the implication of a node over the other. However, to our knowledge, none have explored the contribution of node i's past, over the future state of node j. In this paper, we propose a Spatio-Temporal Graph Mixformer (STGM) network, a highly optimized model with low memory footprint. We address the aforementioned limits by utilizing a novel attention mechanism to capture the correlation between temporal and spatial dependencies. Specifically, we use convolution layers with a variable fields of view for each head to capture long-short term temporal dependency. Additionally, we train an estimator model that express the contribution of a node over the desired prediction. The estimation is fed alongside a distance matrix to the attention mechanism. Meanwhile, we use a gated mechanism and a mixer layer to further select and incorporate the different perspectives. Extensive experiments show that the proposed model enjoys a performance gain compared to the baselines while maintaining the lowest parameter counts.
Keywords:
Traffic forecasting
Spatio-temporal multivariate time series
Graph neural network
Attention mechanism
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W