Robust Trajectory Similarity Learning for Mobile Phone Signaling Data via Contrastive Augmentation

Trajectory similarity measurement serves as a fundamental operation in trajectory analysis. Numerous similarity measures have been proposed for GPS trajectories. However, the large-scale application of GPS trajectories is constrained by the high cost of data collection. In contrast, mobile phone signaling data (MPSD) offers the unique advantages of massive data volume, wide-area coverage, and long-term continuity through its passive collection mechanisms, making it indispensable for large-scale trajectory analysis. Existing similarity measures suffer significant effectiveness degradation when applied to MPSD trajectories due to the inherent data quality issues (i.e. coarse-grained localization at the base station level, low recording frequency, and pervasive signal drift and oscillation noises). To tackle this, we propose MPSDTCL, a MPSD trajectory similarity measure based on contrastive learning. Specifically, we design five trajectory augmentation methods and construct an encoder that integrates trajectory structural and spatial feature learning. Through self-supervised contrastive learning, the encoder can adaptively suppress the interference of local noises and produce embeddings that capture the global semantics of MPSD trajectories. Experiments on two real-world datasets demonstrate MPSDTCL's superior performance over existing baselines in measuring the similarity between MPSD trajectories.