Trajectory-User Linking via Heterogeneous Preference Graph and Dual-Encoder Mutual Distillation

Trajectory User Linking (TUL) is a critical task in spatio-temporal behavior analysis, which focuses on reidentifying users from anonymous mobility traces. While prior approaches have made progress by modeling the sequential patterns and behavioral dynamics of trajectories, they often neglect the heterogeneous semantics embedded in user interactions, such as varying check-in frequencies, transition patterns, and contextual associations. In this work, we propose Heterogeneous Preference Graph and Dual-Encoder Mutual Distillation (HPG-DEMD), a unified framework that combines multi-view trajectory encoding with structural preference modeling. Specifically, a heterogeneous interaction graph is constructed to learn user prototypes using a heterogeneous graph neural network, providing long-term preference priors. In parallel, an LSTM-based encoder captures short-term sequential behaviors, while a Transformer-based encoder models long-range temporal dependencies. The LSTM-based and Transformer-based encoders are jointly trained under a mutual distillation strategy, which encourages representation consistency and improves generalization. Extensive experiments on three real-world datasets demonstrate that our model achieves state-of-the-art performance on the TUL task. HPG-DEMD consistently outperforms stateof-the-art methods, with up to +10.27%Acc@1 on Weeplaces, and +7.89%Acc@1 on Foursquare-TKY, demonstrating strong robustness and scalability. The source code is available at: https://github.com/Ledingburger/HPG-DEMD.