DeepSense: A novel learning mechanism for traffic prediction with taxi GPS traces

The urban road traffic flow condition prediction is a fundamental issue in the intelligent transportation management system. While extracting the high-dimensional, nonlinear and random features of the transportation network is a challenge, which is very useful to improve the accuracy of traffic prediction. In this paper, we propose DeepSense, a novel deep temporal-spatial traffic flow feature learning mechanism, with large scale Taxi GPS traces for traffic prediction. Deep-Sense includes two switchable feature learning approaches. DeepSense exploits a temporal-spatial deep learning approach for traffic flow prediction with the sufficient spatial and temporal taxi GPS traces in dynamic pattern. Meanwhile, Deep-Sense takes advantage of a supplementary temporal sequence segment matching approach with the temporal transformation of traffic flow state for a given road segment when there are not enough traffic traces. Experimental results show that DeepSense can achieve higher prediction accuracy with nearly 5% improvements compared with existing methods.