NWSTAN: a lightweight dynamic spatial–temporal attention network for traffic prediction

Advanced deep learning technology has promoted the development of high-accuracy traffic prediction algorithms. However, existing algorithms struggle to balance model size and predictive accuracy, lack the ability to model the spatial correlation effectively, and require many parameters, resulting in a significantly larger model size. To overcome these challenges, this paper introduces a new solution, the lightweight Nodewalk Spatial–Temporal Attention Network (NWSTAN). NWSTAN uses an improved lightweight nodewalk algorithm based on Node2vec to model the correlations between sensors on different roads dynamically. Additionally, it employs a spatiotemporal feature extraction module that leverages the attention mechanism to extract relevant features from the traffic flow data. Experiments conducted on two real-world datasets show that NWSTAN reduced the mean absolute error (MAE) by 3-5% compared to the baseline model while significantly reducing the number of parameters (by at least 27%). This makes NWSTAN easier to integrate into existing edge computing devices, broadening its potential for real-world applications.