Dynamic Auto-Structuring Graph Neural Network: A Joint Learning Framework for Origin-Destination Demand Prediction

Solving the demand prediction problem is an important part of improving the efficiency and reliability of ride-hailing services. Spatial-temporal graph learning methods have shown potential in modeling the spatial-temporal dependencies of ride-hailing demand data, but most existing studies focus on region-level demand prediction with only a few researchers addressing the problem of origin-destination (OD) demand prediction. In addition, previous spatial-temporal graph learning methods employ pre-defined and rigid graph structures that do not reveal the instinct and dynamic dependencies of ride-hailing demand data. In this paper, we propose a joint learning framework called Dynamic Auto-structuring Graph Neural Network (DAGNN) to address the origin-destination demand prediction problem. We develop a Dynamic Graph Decomposition and Recombination layer (DGDR) to handle both the graph structure and the graph representation learning problems simultaneously, with graph representations learned from a group of trainable and time-aware edge-induced subgraphs. Experimental results show that our proposed model outperforms ten baseline models with two real-world ride-hailing demand datasets and is efficient in structural pattern discovery. Comparing with existing methods, the significant advantage of the proposed method is that it circumvents the difficulties in defining the underlying graph structure of the researched data.