Crowd Flow Prediction by Deep Spatio-Temporal Transfer Learning

Crowd flow prediction is a fundamental urban computing problem. Recently, deep learning has been successfully applied to solve this problem, but it relies on rich historical data. In reality, many cities may suffer from data scarcity issue when their targeted service or infrastructure is new. To overcome this issue, this paper proposes a novel deep spatiotemporal transfer learning framework, called RegionTrans, which can predict future crowd flow in a data-scarce (target) city by transferring knowledge from a data-rich (source) city. Leveraging social network check-ins, RegionTrans first links a region in the target city to certain regions in the source city, expecting that these inter-city region pairs will share similar crowd flow dynamics. Then, we propose a deep spatio-temporal neural network structure, in which a hidden layer is dedicated to keeping the region representation. A source city model is then trained on its rich historical data with this network structure. Finally, we propose a regionbased cross-city transfer learning algorithm to learn the target city model from the source city model by minimizing the hidden representation discrepancy between the inter-city region pairs previously linked by check-ins. With experiments on real crowd flow, RegionTrans can outperform state-ofthe-arts by reducing up to 10.7% prediction error.