Taxi demand forecasting based on the temporal multimodal information fusion graph neural network

Online taxi-hailing service is an essential part of a modern intelligent transport system. Accurate taxi demand forecast can reduce the users' waiting time, improve the taxi utilization rate, and optimize transportation efficiency. However, since taxi demand depends on numerous factors, it is difficult to achieve an accurate forecast using only single modality information. Thus, in this paper, a graph neural network model that combines multimodal information is proposed. The taxi demand forecasting is regarded as a time-series feature-processing task. We take each time step as the node in the graph. The node features are initialized with multimodal information and updated based on a novel message passing mechanism with multimodal attention. Experiments were conducted to compare our proposed method with multiple baseline methods on public datasets, and the experimental results show that our method effectively reduces the forecasting error. Finally, the analysis of the factors influencing the taxi demand forecast is presented.