Graph Convolutional Modules for Traffic Forecasting

Graph convolutional network is a generalization of convolutional network for learning graph-structured data. In some of the recent works on traffic networks, a few graph convolutional blocks have been designed and shown to be useful. In this work, we extend the ideas and provide a systematic way of creating graph convolutional modules. The method consists of designing basic weighted adjacency matrices as the smallest building blocks, defining partition functions that can partition a weighted adjacency matrix into M matrices that can also serve as building blocks, and finally designing graph convolutional modules using the building blocks. We evaluate some of the designed modules by replacing the graph convolutional parts in STGCN and DCRNN, and find 8.4% to 25.0% reduction in speed estimation error.