The paper presents MobiGeaR, a novel framework that generates human mobility data using large language models (LLMs). Traditional methods of collecting human mobility data are expensive and involve privacy risks, creating a need for high-quality generative mobility models. Previous efforts focused on learning behavior distribution from training samples, which failed to capture the coherent intentions driving mobility behavior. MobiGeaR addresses these issues by reformulating mobility generation as a commonsense reasoning problem. The framework uses LLMs to generate mobility behavior based on context-aware prompts. It also employs a divide-and-coordinate mechanism that combines LLM reasoning with a mechanistic gravity model. The approach significantly improves the semantic-awareness of mobility generation, with experiments showing a 62.23% improvement in intention accuracy.
Publication date: 16 Feb 2024
Project Page: https://anonymous.4open.science/r/MobiGeaR-C57C
Paper: https://arxiv.org/pdf/2402.09836
