Machine Unlearning on Trajectory Data: An Experimental Analysis

Machine Learning (ML) models have become essential for trajectory data analysis, supporting tasks such as similarity learning, map matching, simplification, and recovery. While learning from trajectory data enables personalized and location-based services, the ability to unlearn specific samples is equally crucial for removing outdated information, mitigating privacy risks, and complying with emerging data regulations. Existing machine unlearning methods, however, have been developed almost exclusively for the Computer Vision (CV) domain and their applicability to trajectory learning remains largely unexplored. In this work, we present the first systematic study of machine unlearning for trajectory data. We investigate (i) whether and how representative unlearning algorithms can be adapted to trajectory learning models, (ii) how to evaluate their effectiveness, and (iii) how unlearning outcomes vary across datasets, tasks, and unlearning scenarios. To this end, we design a comprehensive experimental framework covering ten unlearning algorithms, four key trajectory learning tasks, and multiple evaluation metrics on three real-world datasets. Our study reveals the unique characteristics and sensitivities of unlearning on trajectory data, provides insights into the limitations of existing approaches, and highlights open challenges for future research. The results establish a foundation for building standardized benchmarks and advancing machine unlearning in trajectory analytics. The code is available at https://github.com/yuanjun-liu/TrajMU.