Temporal Attention-Based Interactive Dynamic Graph Convolutional Network For Traffic Flow Prediction

Accurate traffic flow prediction is crucial for mitigating road congestion. While previous studies have developed successful models for predicting future traffic flow, most of them lack effective modeling of spatiotemporal dependencies and struggle with capturing multi-scale spatiotemporal dependencies. This paper introduces a novel approach called Temporal Attention-Based Interactive Dynamic Graph Convolutional Network (TAIDGCN) to effectively model the spatiotemporal characteristics of traffic flow. Firstly, a feature augmentation layer enhances the interactivity of temporal features. Secondly, an interactive dynamic graph convolutional network captures spatiotemporal features through interactive learning. A dynamic graph convolutional block is designed to capture changing spatial correlations using a novel dynamic graph generation method. Finally, a specialized temporal attention mechanism is constructed to capture dynamic temporal dependencies, enabling accurate long-term traffic flow prediction by utilizing local context and numerical sequence representation transformation. Experimental results on real traffic datasets demonstrate the superior performance of the proposed model compared to existing approaches