Dynamic Deeper Graph Convolutional Network for Traffic Prediction

Abstract

Graph convolutional networks (GCNs) are widely used in traffic prediction because they can better represent the structure of traffic network. GCNs can obtain deeper feature representation by aggregating neighborhood features. However, the network is prone to over-smoothing, as the aggregation neighborhood deepens. GCNs used for traffic prediction usually aggregate neighborhood features of order 2 or 3. To obtain deeper features, we use an adaptive hidden layer connection method to deepen neighborhood aggregation in traffic graph network for the first time. It adaptively adjusts the weight of hidden layer to increase the initial connection and hidden layer connection, which can obtain deeper neighborhood features and alleviate over-smoothing. We evaluated the model using real data from the road network in Beijing, and it showed good performance, especially in the long-term prediction.