.png){kind=icon}
.png){kind=icon}
Return
Cross-dimensional global interactive transformer for traffic forecasting
.png){kind=icon}
PRE
AI
DOI：10.1016/j.dsp.2024.104974.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Traffic forecasting is a crucial element of intelligent transportation system and has numerous practical applications for public safety and resource management. Accurate traffic forecasting presents a significant challenge due to the intricate and ever-changing spatio-temporal correlations of traffic data. Consequently, numerous spatio-temporal networks have been developed to address the challenge. However, many existing spatio-temporal networks are based on end-to-end training and still suffer from some issues: (i) The end-to-end learning may lead to overfitting and capture some spurious correlations of traffic data; (ii) Most networks ignore the cross-dimensional global interactions between different dimensions of traffic data. This paper proposes a new cross-dimensional global interactive Transformer (CDGIT) to address these issues mentioned above. The proposed CDGIT decomposes complex traffic data into fluctuating events and stable trends using the discrete wavelet transform (DWT), and then utilizes a dual-channel spatio-temporal network to capture spatio-temporal dependencies. Specifically, cross-dimensional interactive attention (CDIA) and global multi-scale attention (GMSA) modules are proposed to extract the temporal correlation of events and trends, respectively, and local- global Transformer encoder (LGTE) module is developed to extract local and global spatial correlations. In addition, graph position encoding based on graph wavelets is introduced to balance the local-global structural information. Finally, the forecasting results are obtained by an adaptive fusion module that uses the attention mechanism and is learned by backpropagation. A series of comprehensive experiments conducted on three actual traffic datasets have demonstrated that the proposed CDGIT is more effective and superior compared to other traffic forecasting methods.
Keywords:
Intelligent transportation system
Traffic forecasting
Attention mechanism
Transformer
Position encoding
Journal
IF:
3.6
Papers:
9.9K
Citations:
1.7W