Attention augmented spatial-temporal embedding multilayer perceptrons for traffic flow prediction

Abstract The linear framework has become an alternative paradigm for traffic flow prediction models with more predictive performance and less training cost than Spatial-Temporal Graph Neural Networks (STGNNs). However, challenges such as insufficient learning capacity for spatial-temporal interaction and difficulty in perceiving global contextual information still persist in the linear framework. In this paper, we propose a novel Attention Augmented Spatial-Temporal Embedding Multilayer Perceptrons (ASTMLP) to capture the complex spatial-temporal pattern for traffic flow prediction. We introduce a novel input embedding method by combining it with external attention to capture the global contextualization for the input sample data. Afterward, a dual-stream fusion architecture based on MLP is proposed to capture the spatial-temporal dependence in a decoupling-recoupling manner. Experimental results in four real-world datasets demonstrate that ASTMLP consistently outperforms previous state-of-the-art models, improving all evaluation metrics by approximately 0.07%–2.59%. In particular, compared to STGNNs, ASTMLP can improve performance by 0.55%–7.58%, with a training time reduction from 2.90%–57.01%.