LEARNABLE GRAPH CONVOLUTIONAL NETWORK FOR TRAFFIC FORECASTING APPLICATIONS

Accurate and timely traffic forecasting has been a critically important topic in intelligent transportation systems. Considering the road network topology, various graph convolution network based deep learning models have been proposed. However, existing methods have insufficient interpretability and weak capability when modelling the heterogeneous spatial dependence. In this paper, a multi-temporal learnable graph convolution network (MT-LGC) model is proposed to solve the network-based traffic forecasting problem. The MT-LGC consists of three independent components. Each component integrates a learnable graph convolution (LGC) network for learning the heterogenous spatial dependence and a long short-term memory (LSTM) network for temporal dependence modelling. Experiments on two real-world traffic datasets show that our model MT-LGC outperforms the state-of-the-art baselines. Results of case studies also show that the proposed LGC can be integrated into these state-of-the-art methods as a general graph convolution operator and can achieve higher forecasting accuracy.