Mobility Anomaly Generation using LLM-Driven Behavior with Kinematic Constraints

The paper presents a novel end-to-end generative framework for synthesizing human trajectory anomalies, addressing the lack of annotated datasets for such events. It utilizes Large Language Model (LLM) agents to introduce semantically meaningful behavioral anomalies into baseline simulated trajectories, while employing map-constrained routing reconstruction to maintain spatial validity and a context-aware spatial noise model to reflect real-world GPS sensor degradation. This approach is significant for practitioners as it enables the creation of realistic, annotated datasets for training and testing anomaly detection systems in mobility data, thereby enhancing the robustness of spatial data mining applications.