.png){kind=icon}
.png){kind=icon}
Return
Transformer network with decoupled spatial-temporal embedding for traffic flow forecasting
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-023-05126-x.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Over the past few years, there has been significant research on applying Transformer models to time series prediction, yielding promising results. Simultaneously, researchers have begun exploring the utilization of Transformers for traffic prediction in order to mitigate the nonlinear spatial-temporal correlation inherent in traffic data. Some of these studies have attempted to characterize spatial-temporal features by incorporating embedding structures, with the goal of improving performance of the model. However, existing methods have not adequately addressed the issue of spatial-temporal correlation. To address these limitations, we propose the Transformer Network with Decoupled Spatial-Temporal Embedding (DSTET) model for traffic flow prediction. The key aspect of our model is its ability to effectively decouple the spatial and temporal embedding through the implementation of the Decoupled Spatial-Temporal Embedding structure. This structure enhances the characterization of spatial-temporal features, ultimately improving the performance of traffic prediction based on the Transformer model. Through experiments conducted on six real-world traffic datasets, our model consistently outperforms multiple baseline models, demonstrating its capability to address the identified problems. Moreover, we substantiate the efficacy of the suggested components via ablation experiments and furnish a thorough analysis of the attention weight matrix to clarify the functioning of the model.
Keywords:
Transformer
Traffic Forecasting
Embedding
Journal
IF:
3.5
Papers:
7.5K
Citations:
1.7W