Trajectory-user linking via supervised encoding

Abstract

With the explosive growth of Location-Based Services (LBS), a substantial amount of geolocation data, containing end-user private information, is amassed, posing severe privacy risks. Trajectory-User Linking (TUL) is a trajectory mining task aimed at linking trajectories to their generators. Recent research has introduced deep learning-based TUL models. However, these models face challenges related to limited data quality and inadequate extraction of bidirectional and multi-topic semantic information from trajectories. In this study, we propose Trajectory-User Linking via Supervised Encoding (TULSE), centered on supervised encoding of location points and trajectories to address the TUL task. Specifically, TULSE extracts spatial and temporal information from location points through a novel method named Supervised Spatiotemporal Encoding. Additionally, TULSE employs a BiLSTM with multi-head attention to capture bidirectional and multi-topic semantics from trajectories. Furthermore, recognizing the limitations of current evaluation metrics, we introduce a novel metric named Hierarchical Privacy Loss (HPL). HPL offers a more detailed assessment of TUL solutions by statistically analyzing the distribution of prediction accuracy among users. We conduct extensive experiments on two benchmark datasets, and empirical results show that TULSE outperforms existing TUL methods.

Keywords

• trajectory-user linking
• deep learning
• location embedding
• attention mechanism
• privacy risk measurement