Probabilistic Bayesian learning with long-tail awareness for trajectory-user linking.

Abstract

Recent location-based social networks (LBSN) provide us with a wealth of user check-in trajectories, offering researchers more prominent opportunities to develop effective practices for understanding the mobility patterns of different groups. In particular, the task of Trajectory-User Linking (TUL), which aims to accurately associate unlabeled check-in trajectories with the users who generated them, has received significant attention in the GeoAI community. Prior investigations have shown that the sparsity of check-ins hinders the capture of diverse user interests, requiring massive efforts to handle this concern. However, the diversity and individual preferences inevitably lead to a more thorny issue regarding the imbalanced collection of historical check-ins by various users, commonly referred to as the long tail phenomenon. In particular, the disparity in check-ins among various users intensifies the inference preferences towards the head users while overlooking those who contribute less often (i.e., tail users). That is, such an imbalance eventually leads to the mislabeling of unlinked trajectories as the head users, diminishing the likelihood of correctly classifying the tail users. In this study, we propose a novel probabilistic Bayesian learning solution, called LongTUL, primarily addressing the concern of the long tail behind the TUL task. In LongTUL, we build a Check-in Engagement Compromise (CEC) mechanism to harmonize the participation levels of various users before the training phase. Then, we are motivated by recent variational Bayes and develop a Probabilistic Trajectory Learning (PTL) procedure to encode each trajectory into a latent space. In contrast to the prior arts, we involve the Laplacian approximation regarding the latent representations to address the amortization errors as well as alleviate the effects of long-tail characteristics. Furthermore, we develop a reweighted classifier designed to ensure equitable inference between head users and tail users. Finally, our experiments conducted on three real-world check-in datasets demonstrate the superiority of LongTUL against representative TUL solutions, especially in addressing the long-tail issue.