A User Trajectory Identification Model with Fusion of Spatio-Temporal Behavior and Social Relation
ZHANG Wei,LI Yang,ZHANG Ji,WANG Jian-Yong
DOI: https://doi.org/10.11897/SP.J.1016.2021.02173
2021-01-01
Chinese Journal of Computers
Abstract:With the flourish of location-aware online social platforms, user behavior of mobility has been greatly enriched, which promotes the relevant studies on the user trajectory identification problem. For a given target trajectory, this problem aims to identify a specific user that the trajectory belongs to, which is 2174 计 算 机 学 报 2021 年 beneficial for understanding the mobility patterns behind user trajectories and could provide positive influence on a variety of downstream applications, such as personalized recommendation, to name a few. By far, a few existing relevant studies try to utilize multi-class classification methods to tackle the problem. However, these studies still face two main unresolved challenges which need to be addressed in the literature: the sparsity of user trajectories and the large number of categories to be used for classification. Among the two challenges, the reason of the first one is that users tend to choose only a limited number of visited locations to be published in online social medias and their preferred visited POIs are distributed in local regions; the second challenge is caused by the fact that in multi-class classification, each category denotes one user and the number of users is large. To address the above two challenges, this paper aims to effectively utilize the two types of information, i. e., sequences of timestamps in user trajectories and social relations among users, both of which have not been investigated by previous studies for the considered problem. On the one hand, timestamps could be used as additional information to alleviate the sparsity issue. On the other hand, social relations could be leveraged to characterize the correlations between users, which in turn helps user representation learning. To effectively leverage the two types of information for the studied problem, we propose a novel model which couples Neural Temporal Point Process with Graph Neural Network, named NTPP-GNN. This model composes of three modules for the spatial aspect, the temporal aspect, and the social aspect, respectively. In the spatial module, bidirectional recurrent neural networks are exploited to characterize the sequential nature of visited locations contained in trajectories. In the temporal module, we propose a bidirectional neural temporal point process to capture the continuity of time in both forward and backward directions, which could be further employed for facilitating the temporal representation learning of trajectories. In the social relation module, graph neural networks are exploited to encode user social relations into their representations. The whole model NTPP-GNN adopts an end-to-end learning fashion to optimize the above three modules jointly. This could ensure that each learnable module to be compatible with each other for achieving better trajectory identification performance. To validate the effectiveness of NTPP-GNN, we conduct extensive experiments on three widely used available datasets(i. e., Foursquare, Gowalla, and Brightkite). The experimental results indicate that: (1)the proposed NTPP-GNN improves the best-performed baseline w. r. t. ACC@1 by 7.0% in average; (2)each module in NTPP-GNN indeed makes a positive contribution to the final multi-class classification performance; and(3)the developed bidirectional neural temporal point process brings additional performance gains over conventional neural temporal point process approaches that only consider the forward direction in modeling sequential timestamps.