DUTD: A Deeper Understanding of Trajectory Data for User Identity Linkage

Abstract

The widespread adoption of mobile devices and sensing technologies has significantly boosted the availability of trajectory data, thereby facilitating studies on users’ spatio-temporal behaviors. Among these studies, the task of trajectory-based user identity linkage (UIL) has attracted extensive attention recently. Existing efforts for the task mainly rely on data mining methods and sequence models to extract features and measure user identity similarity. Most of them commonly regard trajectories as mere sequences, thus failing to take full advantage of the hidden spatio-temporal characteristics within trajectories, such as temporal regularities and mobility patterns. To fill the gap, we propose a novel model namely DUTD (A Deeper Understanding of Trajectory Data for User Identity Linkage). Specifically, we first adopt the skip-gram module on historical trajectories to generate grid representations of each trajectory. Subsequently, a transformer-based encoder is designed to capture long-term temporal regularities and mobility patterns. Particularly, to enhance the effectiveness and efficiency of the encoder, we introduce two self-supervised learning tasks: trajectory recovery and trajectory contrastive learning. Ultimately, to capture the inter-trajectory correlations between users, a matcher comprising a multi-interaction module and a prediction layer is designed. The extensive experiments conducted on two real-world datasets demonstrate the superiority of our proposed model compared to the state-of-the-art methods.