Long-Short-Term Expert Attention Neural Networks for Traffic Flow Prediction

Accurate traffic flow prediction is crucial for Intelligent Transportation Systems (ITS). Traffic patterns are influenced by both temporal dynamics and road network structure, requiring the extraction and utilization of spatial and temporal features for effective prediction. In long-term prediction tasks, spatial features lose importance, emphasizing the need for a nuanced approach that prioritizes temporal features or historical data across various scales. Previous studies have employed uniform models for both short-term and long-term tasks, potentially limiting the model's expressive capacity. To address this, we propose the Long-Short-Term Expert Attention Neural Networks (LSTANN) tailored to diverse prediction tasks. In this model, the short-term expert network predominantly extracts the dynamic spatial and temporal features of current traffic flow, facilitating short-term traffic pattern predictions. Simultaneously, the long-term expert network focuses on capturing the temporal characteristics of traffic flow across varying historical scales. The Mixture of Expert (MoE) component adeptly processes data at different scales, enabling seamless adaptation to diverse expert networks. Comprehensive experimental results substantiate the superior performance of the proposed model compared to other baseline methods.