A Trajectory-Based User Movement Pattern Similarity Measure for User Identification

Recently, matching the cross-site user accounts based on user trajectory similarity has been attracting much attention, which benefits many applications. Most of existing works measure the user trajectory similarity based on the proximity of stay points. However, since the user trajectory is noisy, sparse and biased, the trajectory similarity measure is not a trivial task. In theory, the user movement pattern inherent in user trajectory is unique to the user and site-independent, which may be the key to address the above question. Thus, we propose a trajectory-based user movement pattern similarity measure for user identification. Specifically, we first find the frequent user stay points at each time slot to represent the user daily activities, and then present a similarity measure method for user daily activities, where we assign a global popularity and local popularity to stay point with the aim of accurately characterizing the contribution of stay point to the similarity. To represent the user-specific and site-independent movement pattern, we propose a trajectory-oriented embedding method, called T-LINE, which preserves the similarity of user daily activities. Finally, based on the similarity of movement patterns between a user and his/her candidates, we select top $n$ similar users as the matching users. The experiments conducted on three ground-truth datasets show a significant improvement compared to the representative works.