Multi-Attribute Spatial-temporal Graph Convolutional Network for Taxi Demand Forecasting

Abstract

Accurate forecasting taxi demand help reduce waiting time for drivers and passengers as well as ease traffic congestion. However, most of the current research work has mostly ignored the impact of historical cab inflows and potential spatial dependencies between different regions on taxi demand. In view of this, this paper integrates several attributes affecting taxi demand and develops a multi-attribute spatial-temporal graphical convolutional network model (MASTGCN) with the expectation of accurately predicting the MASTGCN model is designed to accurately predict the demand for rental cars. Specifically, the MASTGCN model is designed with four components, which model the temporal dependence of taxi demand on the demand series at the near moment, the daily demand series, the historical taxi inflow series, and the daily demand series, respectively. The components are designed to model the temporal dependence of taxi demand on proximity demand series, daily demand series, historical cab inflow series, and the potential spatial dependence among different regions. To demonstrate the effectiveness of the MASTGCN model, we compare it with five benchmark models commonly used for traffic forecasting and three metrics, RMSE, MAE and MAPE, are used for evaluation. The experimental results show that the MASTGCN model, which incorporates multiple attributes, can more accurately the multiattribute MASTGCN model can predict taxi demand more accurately.