Decoding digital footprints: user re-identification through mobility pattern decomposition and collaborative fusion

Modeling the complexity of human mobility from vast trajectory data presents a core scientific challenge. In the task of user re-identification, performance is constrained by modeling paradigms that fail to resolve the inherent heterogeneity between short-term dynamics and long-term habits. Prevailing monolithic approaches treat mobility as a homogeneous sequence, while even recent multi-branch models often lack principled decomposition or effective synergistic fusion. In this paper, we argue that addressing this fundamental limitation requires a paradigm shift from monolithic modeling to a philosophy of “decomposition-and-synergy.” Our proposed framework instantiates this philosophy through a parallel, dual-branch architecture: one branch leverages a temporal convolutional network (TCN) to decompose micro-level causal dependencies, while a concurrent Transformer branch distills macro-level periodic paradigms. Crucially, we design a Co-Attention mechanism that establishes a bidirectional dialogue between these heterogeneous representations to enable synergistic reasoning, culminating in a more holistic and discriminative user profile. Extensive experiments on three real-world LBSN datasets establish a new state-of-the-art. Our full dual-branch model yields results that demonstrate substantial improvements. On the Foursquare (259-user) dataset, our model achieves an ACC@1 of 60.27% and a Macro-F1 of 60.24%, outperforming the strongest baseline, DeepTUL, by 3.11 percentage points. On the challenging Brightkite (259-user) dataset, our model’s ACC@1 of 63.21% constitutes a 2.57 percentage point gain over the same baseline. These results provide compelling empirical evidence for the superiority of our proposed decomposition-and-synergy paradigm.