Trajectory representation learning based on road network partition for similarity computation

In the tasks of location-based services and vehicle trajectory mining, trajectory similarity computation is the fundamental operation and affects both the efficiency and effectiveness of the downstream applications. Existing trajectory representation learning works either use grids to cluster trajectory points or require external information such as road network types, which is not good enough in terms of query accuracy and applicable scenarios. In this paper, we propose a novel partition-based representation learning framework PT2vec for similarity computation by exploiting the underlying road segments without extra information. To reduce the number of words and ensure that two spatially similar trajectories have embeddings closely located in the latent feature space, we partition the network into multiple sub-networks where each is represented by a word. Then we adopt the GRU-based seq2seq model for word embedding, and a loss function is designed based on spatial features and topological constraints to improve the accuracy of representation and speed up model training. Furthermore, a hierarchical tree index PT-Gtree is built to store trajectories for further improving query efficiency based on the proposed pruning strategy. Experiments show that our method is both more accurate and efficient than the state-of-the-art solutions.