CRII: III: Unlearning Spatial-Temporal Imitation Learning for Trustworthy Urban Intelligence

About This Grant

Artificial intelligence (AI) systems are increasingly used to understand and replicate human decision-making in urban environments, from ride-sharing and delivery to traffic planning. These systems rely on spatial-temporal imitation learning (STIL), where models learn from human mobility patterns to optimize services. However, as these models ingest massive amounts of sensitive user data (e.g., GPS traces), new challenges arise in ensuring privacy and compliance with emerging “right-to-be-forgotten” regulations. This project will develop new technologies that enable STIL models to selectively forget specific user data without retraining from scratch. The novelty of this project lies in its ability to effectively erase individual trajectories or behaviors embedded in complex user data, while preserving the utility and accuracy of the original models. The proposed technologies will empower users with control over their data and help ensure that AI systems used in urban intelligence are not only effective but also privacy and legally compliant. This project will pioneer a new research direction in machine unlearning within the spatial-temporal imitation learning (STIL) domain by introducing the SIFT (Spatial-temporal Imitation ForgeTting) framework. SIFT is designed to address three key challenges: (1) overlapping and dispersed trajectories that make precise data unlearning difficult, (2) spatial-temporal heterogeneity and sparsity that can degrade model performance when unlearning data from data-sparse regions, and (3) lack of theoretical guarantees for effective unlearning. The research will develop innovative methods including negation trajectory generation and reward shaping to remove privacy-sensitive data, spatial-temporal partitioning and similarity-based data augmentation to handle data sparsity and variability, and differential privacy techniques to ensure provable unlearning guarantees. These contributions will be evaluated using large-scale, real-world datasets from ride-hailing and transit systems. The project will also integrate unlearning research into undergraduate and graduate courses. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.