Foreseeing private car transfer between urban regions with multiple graph-based generative adversarial networks

Private car transfer indicates that people drive private cars and travel between urban regions to perform daily activities. Foreseeing private car transfer between urban regions can facilitate a broad scope of applications ranging from route planning, hot region discovery to urban computing. However, three challenges remain. i ) Private car transfer between regions is affected by multiple spatio-temporal correlations. ii ) Transfer records are highly sparse and imbalanced. iii ) Modeling the stay duration of private cars. In this paper, we model private cars’ travel in urban regions as the spatio-temporal graph and formulate private car transfer foreseeing as the time-evolving adjacency matrix prediction of the graph. To specify, we propose MG-GAN ( M ultiple G raph-based G enerative A dversarial N etwork) to predict private car transfer. For one thing, we design multi-graph dense convolutions with gated recurrent networks as the generative network to capture multiple spatio-temporal correlations. For another, the attentive multi-graph convolutional network is designed as the discriminative network to learn the stay duration correlations of private cars in each region. The iterative adversarial processes between generating and discriminating networks enhance the MG-GAN’s ability to tackle the sparse data problem. Besides, a topic clustering algorithm based on multi-source data fusion is proposed to balance the fused data. Extensive experiments on the real-world private car and taxi trip datasets demonstrate that MG-GAN performs better than the state-of-the-art baselines.