Knowledge Graph Enhanced Dynamic Multi-Graph Convolutional Network for Traffic Origin-Destination Forecasting

Traffic origin-destination forecasting is a crucial yet formidable task in intelligent transportation systems, due to the complicated spatial-temporal dependencies. However, existing spatial-temporal modelling methods are either solely reliant on data or excessively dependent on domain knowledge, and few researchers consider both of them in combination. In this paper, a knowledge graph enhanced dynamic multi-graph convolutional network is proposed for traffic origin-destination forecasting, which integrates a knowledge graph embedding method to explore the latent patterns of traffic flows. Specifically, considering spatial-temporal dependencies modelling, a traffic demand knowledge graph is constructed and pre-trained by a translation-based model to obtain the spatial-temporal embeddings, then a knowledge-enhanced spatial-temporal dynamic graph is constructed based on the pre-trained knowledge graph embeddings. Besides, regarding origin-destination correlation modelling, a similarity-based dynamic graph is designed to capture the dependencies of both origins and destinations. Subsequently, a multi-branch graph convolutional framework is developed with a sequence of RNN followed with 2DGCN to capture the spatial-temporal dependencies of the two beforementioned graphs. Furthermore, a step-wise decoder and an average ensemble strategy are used to obtain the final multi-step forecasting results. Experimental results demonstrate the state-of-the-art performance of our proposed model in leveraging complex spatial-temporal dependencies.