With well-defined graphs, graph convolution based spatiotemporal neural networks for traffic prediction have achieved great performance in numerous tasks. Compared to other methods, the networks can exploit the latent spatial dependencies between nodes according to the djacency
relationship. However, as the topological structure of the real road network tends to be intricate, it is difficult to accurately quantify the correlations between nodes in advance. In this paper, we propose a graph convolutional network based adaptive graph learning algorithm (AdapGL) to acquire the complex dependencies. First, by developing a novel graph learning module, more possible correlations between nodes can be adaptively captured during training. Second, inspired by the expectation maximization (EM) algorithm, the parameters of the prediction network module and the graph learning module are optimized by alternate training. An elaborate loss function is leveraged for graph learning to ensure the sparsity of the generated affinity matrix. In this way, the expectation maximization of one part can be realized under the condition that the other part is the best estimate. Finally, the graph structure is updated by a weighted sum approach. The proposed algorithm can be applied to most graph convolution based networks for traffic forecast. Experimental results demonstrated that our method can not only further improve the accuracy of traffic prediction, but also effectively exploit the hidden correlations of the nodes. The source
code is available at https://github.com/goaheand/AdapGL-pytorch.