Dynamic meta-graph convolutional recurrent network for heterogeneous spatiotemporal graph forecasting

Abstract

Spatiotemporal Graph (STG) forecasting is an essential task within the realm of spatiotemporal data mining and urban computing. Over the past few years, Spatiotemporal Graph Neural Networks (STGNNs) have gained significant attention as promising solutions for STG forecasting. However, existing methods often overlook two issues: the dynamic spatial dependencies of urban networks and the heterogeneity of urban spatiotemporal data. In this paper, we propose a novel framework for STG learning called Dynamic Meta-Graph Convolutional Recurrent Network (DMetaGCRN), which effectively tackles both challenges. Specifically, we first build a meta-graph generator to dynamically generate graph structures, which integrates various dynamic features, including input sensor signals and their historical trends, periodic information (timestamp embeddings), and meta-node embeddings. Among them, a memory network is used to guide the learning of meta-node embeddings. The meta-graph generation process enables the model to simulate the dynamic spatial dependencies of urban networks and capture data heterogeneity. Then, we design a Dynamic Meta-Graph Convolutional Recurrent Unit (DMetaGCRU) to simultaneously model spatial and temporal dependencies. Finally, we formulate the proposed DMetaGCRN in an encoder–decoder architecture built upon DMetaGCRU and meta-graph generator components. Extensive experiments on four real-world urban spatiotemporal datasets validate that the proposed DMetaGCRN framework outperforms state-of-the-art approaches.