GHTPAN:A Graph Head-Tail Partner Attention Network for Traffic Flow Prediction

Traffic prediction is the core of intelligent transportation system (ITS). However, it is a challenging task to construct an accurate prediction model that can cope with the complexity of road network in the real world. In the traffic network, the traffic flow of different road sections or intersections changes dynamically with time, and there is mutual influence between the flow of adjacent areas in space. Most of the existing traffic flow prediction methods lack the ability to model the dynamic spatio-temporal correlation in traffic data, so they cannot obtain satisfactory prediction results. In order to obtain nonlinear traffic flow data and avoid the problem of gradient disappearance or explosion when processing time series data. We propose an innovative approach called Graph Head-Tail Partner Attention Network (GHTPAN). Specifically, the head-tail partner attention module is introduced. By employing a dual mechanism that encompasses channel attention and weighted distribution, the system intelligently identifies and assigns appropriate weights to the sensor nodes within the road network. This intelligent mechanism not only focuses on adjacent nodes but also dynamically captures the spatial interactions between non-adjacent nodes. Combined with the graph convolutional neural network, the local spatial dynamic correlation characteristics of the road network are extracted.A series of comprehensive experiments using highway traffic datasets PEMS03, PEMS04 and PEMS08 not only verify the excellent performance of the GHTPAN model, but also highlight its cutting-edge capabilities.