Bidirectional Multi-grain Graph Convolution Network for Origin-Destination Demand Prediction

Effective origin-destination (OD) demand forecasting is crucial for optimizing travel planning and enhancing the efficiency of taxi and car-hailing services by minimizing passenger waiting time. Traditional OD demand studies often overlook the bidirectional nature of travel patterns, treating connections as one-way links, which can compromise prediction accuracy. Recognizing the interdependent and distinct characteristics of travel demands between two points, this research introduces a novel bidirectional multi-grain graph convolution network (BiDP) for more precise OD demand prediction. BiDP accounts for the bidirectional relationship by calculating weight distributions in both directions, from origin to destination and vice versa, and integrating these to determine a bidirectional relevance level. To tackle the sparsity of OD data, the model employs a multi-grain feature extraction technique that leverages dual-scale convolutional layers to extract pertinent demand information. The integration of a graph convolution network and a global temporal correlation module within BiDP enables the capture of broader spatiotemporal dependencies, thereby refining the prediction process. Comprehensive experiments conducted on two real-world datasets have conclusively shown that BiDP surpasses current leading methodologies in terms of predictive precision, thereby marking a noteworthy progression within the domain of transportation demand estimation.