STTraj2Vec: A Spatio-Temporal Trajectory Representation Learning Approach
Jiahui Zhu,Xinzheng Niu,Fan Li,Yixuan Wang,Philippe Fournier-Viger,Kun She
DOI: https://doi.org/10.1016/j.knosys.2024.112207
IF: 8.139
2024-01-01
Knowledge-Based Systems
Abstract:Computing trajectory similarity plays a critical role in various spatio-temporal applications that involve trajectory analysis. In recent years, trajectory representation learning has been extensively studied and applied for trajectory similarity calculation. However the majority of existing algorithms for trajectory representation generally have two problems. The first problem is the emphasis of spatial similarity over temporal similarity, and even to discard the temporal dimension of spatio-temporal trajectories. As a result, the outputs of these approaches cannot fully represent the similarity of spatio-temporal trajectories. The second problem is the introduction of additional information, such as the topology of the road network, which increases the uncertainty of capturing the spatio-temporal correlation of trajectories and prevents their application in scenarios where it is difficult to obtain such information. This poses a significant challenge when dealing with complex and time-varying traffic networks. This paper proposes a novel method, named STTraj2Vec (Spatio-temporal Trajectory 2 Vector), which relies only on spatio-temporal trajectories to capture their similarity without spatio-temporal separation. This takes into account the whole spatio-temporal trajectory information. In this method, an extended clustering algorithm is introduced, which maps the trajectory into a point-region quadtree, and constructs a time-varying virtual network structure based on the point-region quadtree. In this method, an extended clustering algorithm is introduced, which maps each trajectory into a point-region quadtree, and then completes density clustering through the adjacency relation of leaf nodes to construct a time-varying virtual network structure. This virtual network structure not only considers the spatial proximity, but also the time, so as to reflect the spatio-temporal characteristics of the trajectory more accurately. Then, a novel spatially and temporally integrated random walk algorithm is designed, which carries out spatiotemporal random walk on the virtual network structure, to capture the spatiotemporal characteristics of the trajectory, and thus obtains the representation of all nodes in the virtual network. Furthermore, each trajectory is converted into a sequence of vectors on the virtual road network according to the latitude, longitude, and time of the trajectory points. Finally, based on these node representations and trajectories, a transformer model with ranking loss is employed to capture the distinct contributions of the various locations and times to the similarity computation and encode each sequence of vectors into target vectors. Experiments on two public datasets show that STTraj2Vec is superior to the state-of-the-art methods in terms of effectiveness for top-k trajectory similarity search and trajectory clustering, while exhibiting low parameter sensitivity and high model robustness.