Local-Perception-Enhanced Spatial–Temporal Evolving Graph Transformer Network: Citywide Demand Prediction of Taxi and Ride-Hailing

Abstract

Accurate prediction of demand for traditional taxi and ride-hailing services is crucial for addressing supply-demand imbalances. However, recent studies based on global adaptive graphs, local spatial-temporal graphs, and self-attention mechanisms struggle to effectively capture the dynamic and intricate relations in demand. Moreover, existing dynamic graph generators face challenges in efficiently producing high-quality graphs to learn the diverse interactions among zones along time axis and their shared patterns spanning various time scales. To solve these challenges, we propose a novel Local-Perception-Enhanced Spatial-Temporal Evolving Graph Transformer Network (LPE-STGTN), aimed at improving the effectiveness and efficiency of extracting intricate local dependencies in taxi demand. Specifically, we elaborately design a spatial-temporal evolving graph generator to absorb shared and diversified inter-zone relations across different temporal periodicities and specific interactions among zones within each time step. Furthermore, an attention free transformer with local context (AFT-local) is introduced to effectively learn the correlations between adjacent time steps. Extensive experiments on three taxi datasets of New York and Beijing are carried out to evaluate the superior performance of our model. Compared with the most competitive baseline, our model achieves a balance between effectiveness and efficiency on three datasets, with average training time reduction of 70.66% and average performance improvement of 1.96%.