Novel Trajectory Representation Learning Method and Its Application to Trajectory-User Linking

With the widely used mobile phones, the user's trajectory data can be easily collected by the base station. The trajectory representation learning is the upstream task of trajectory modeling in the transportation systems. And excellent pretraining features will greatly improve the downstream model. However, the trajectory data collected by the base station has a lot of missing values, and the trajectory data contains the coordinate internal features and coordinate context features resulting in the traditional embedding methods in the real space being not suitable for the trajectory data. Therefore, this article proposes novel trajectory representation learning method in the complex space based on the Skipgram (TRMC+Skipgram). The trajectory context is represented by a graph to solve the irregular sampling, where the nodes of the graph are coordinates and the weights of the edges are proportional to the number of co-occurrences of the two nodes. Meanwhile, the coordinates are embedded into the complex space for representing both the coordinate context feature and coordinate internal feature. Through the linking prediction evaluated, the TRMC+Skipgram achieves the state-of-the-art results and yields a 3.7% improvement compared with the best baseline on the AUC metric. In order to verify that the robust pretraining vectors can improve the effect of downstream tasks, the trajectory-user linking model (TULM) is constructed based on TRMC+Skipgram, namely TULM+TRMC+Skipgram. In the task of user-trajectory linking, the TULM+TRMC+Skipgram achieves the state-of-the-art result and yields 1.7% improvement compared with best baseline on the macro-F-1 metric.