O2D: An uncooperative taxi-passenger’s destination predication system via deep neural networks

Predicting passenger’s destination with the partial GPS trajectory is a challenging yet meaningful issue in the taxi industry. Existing destination prediction studies mainly focus on the trajectory data mining algorithms. Moreover, most of them are still unsuitable to be directly applied in real scenarios, due to the problems of unknown trajectory completion degree and unsatisfying prediction performance. In this paper, we present a new destination prediction system with the framework of green edge computing. The system does not require any passengers’ cooperative efforts or privacy information, and can consistently make relatively accurate prediction. Specifically, we propose a LSTM based neural network to automatically estimate the completion degree of partial GPS trajectory. Additionally, to improve the overall performance of destination prediction, we extend our previous deep model to adaptively output multi-granularity prediction results (destination prediction from orientation to destination, i.e., O2D). At last, we deploy the O2D system in real environment and further exploit a real-time recommendation service in a cloud-edge collaboration fashion. Extensive experimental results demonstrate the effectiveness and energy efficiency of our uncooperative passenger’s destination prediction system.