C2TTE: Cross-city Transfer Based Model for Travel Time Estimation

Abstract

Travel time estimation (TTE) is of great importance to many traffic related applications, such as traffic monitoring, route planning and ridesharing. Existing approaches mainly utilize deep neural networks to achieve accurate travel time estimation. These models usually require large-scale trajectory data of the target city, while in reality, it is always difficult for the service providers to obtain sufficient trajectory data in the target city for training the model. To deal with this problem, we propose a cross-city knowledge transfer based travel time estimation solution, namely C2TTE, which is the first to address TTE task via transfer learning. The C2TTE models the travel time in spatial grid granularity first, using not only the sparse trajectory data in the target city, but also the knowledge of data-rich source city via transfer learning. After matching the spatial grid in target city to a suitable grid in source city, we fuse the knowledge between target city and source city to balance their importance adapting to the contextual conditions. In addition, we further use sequential model to represent the path in spatial grid granularity, so that the travel time of the path can be accurately estimated. Comprehensive experiments on two real trajectory datasets have been conducted to demonstrate the superior performance of C2TTE over existing approaches.