An interpretable and efficient multi-scale spatio-temporal neural network for traffic flow forecasting

Existing deep learning based traffic flow forecasting models can effectively learn the complex spatio-temporal dependencies in the traffic network and have become the most widely used traffic flow forecasting architecture in recent years. However, these models have two significant challenges: (1) most of them ignore multi-scale temporal characteristics in traffic sequences; (2) they lack interpretability of traffic flow forecasting. To address the above issues, we propose an interpretable spatio-temporal traffic flow forecasting model with Multi-scale Spatio-Temporal Neural Networks, named MSTNN. Specifically, It first divides original traffic sequences into several patches with different scales to preserve diverse temporal features. Secondly, due to the Kolmogorov Arnold Networks (KAN) have stronger interpretability than existing neural networks by using nonlinear parameter matrix, we design two kinds of advanced variants of KANs in MSTNN, namely Spatial Attention aware KAN (SA-KAN) and Temporal Channel Mixed KAN (TCM-KAN), to enable them to capture spatial structure features and temporal sequence features in traffic data respectively while enhancing the forecasting interpretability. Finally, a fusion module is proposed to aggregate multi-scale temporal features to preserve the multi-scale information. Extensive experiments are conducted to validate the effectiveness of MSTNN and it achieves performance improvement ranging from 0.11 % to 12.65 % on three metrics. The results prove that our MSTNN is effective and interpretable.