Spatiotemporal Multi-Graph Convolutional Network for Taxi Demand Prediction

Abstract

Taxi demand prediction plays an important role in ITS (Intelligent Transportation System). This task is challenging due to the complex spatiotemporal correlations and semantic trends between different locations. Existing work tried to solve this problem by exploiting a variety of spatiotemporal models based on deep learning. However, we observe that more semantic pair-wise correlations among possibly distant roads are also critical for taxi demand prediction. To combine the spatiotemporal correlations with semantic correlations in the traffic network, this paper proposed an end-to-end framework called DeepTDP. First, we defined five kinds of spatial and semantic correlations, which are modeled into multi location graphs and fused by multi-graph convolutional network. Second, LSTM in encoder-decoder network is utilized to capture temporal correlation between future taxi demand values. Besides, a cross-entropy loss function based on error correction is designed to generate taxi demand predictions. Third, we apply a word embedding technique to reduce the dimension of decoded vector in output layer. Finally, we evaluate DeepTDP on two real world traffic datasets, the experiment results demonstrate effectiveness of our approach in comparison with variants of self and other baselines.