Toward Discriminating and Synthesizing Motion Traces Using Deep Probabilistic Generative Models.

Area of Science:

• Data Science
• Machine Learning
• Human Mobility Analysis

Background:

• Trajectory-user linking (TUL) is crucial for location-based services (LBSs).
• TUL exacerbates privacy risks like deanonymization and location recovery.
• Existing methods struggle with disentangled representation and privacy preservation.

Purpose of the Study:

• To develop a semisupervised deep probabilistic framework for joint representation learning and location recovery.
• To create a model that disentangles latent aspects of human trajectories for interpretability.
• To generate synthetic trajectories for privacy protection while retaining mobility information.

Main Methods:

• A novel Semisupervised Trajectory- User Linking model with Interpretable representation and Gaussian mixture prior (STULIG) framework.
• Jointly learning disentangled representations of user trajectories in a semisupervised manner.
• Utilizing a deep probabilistic approach with Gaussian mixture priors for enhanced modeling.

Main Results:

• STULIG effectively learns disentangled representations of human mobility.
• The model demonstrates superior performance in discriminating user traces compared to state-of-the-art methods.
• Synthetic trajectory generation successfully protects user privacy while preserving data utility.

Conclusions:

• STULIG offers an interpretable approach to mobility mining and privacy defense.
• The framework advances TUL by enabling privacy-preserving analysis of human movement data.
• This work provides a robust solution for balancing LBS functionality with user privacy concerns.