KF-TGCN: An Approach to Integrate Expert Knowledge with Graph Convolutional Network for Traffic Prediction

Abstract

There has been continuous research efforts on improving traffic prediction accuracy by exploring the complicated and dynamic spatio-temporal correlation among road network nodes, and it is a challenging problem especially for midto long-term traffic prediction when limited historical data is available. This study proposes a novel knowledge fusion temporal graph convolutional network (KF-TGCN) model to integrate expert knowledge with inherent spatio-temporal correlation which is captured by temporalgraph convolutional network (T-GCN). Our KF-TGCN model has been employed to predict the traffic flow of California highway. Experiment results show that the KF-TGCN model is capable to provide more promising prediction results and significantly reduce the computational complexity comparing to existing models such as gated recurrent units (GRU) and T-GCN.