HMP-LLM: Human Mobility Prediction Based on Pre-trained Large Language Models

Human mobility prediction is fundamental to designing intelligent cities and forecasting how people move during disasters and health emergencies, specifically in emergency responses. However, the existing method for human mobility prediction is mainly designed for normal scenarios and is incapable of handling the case of a pandemic. The COVID-19 pandemic has presented challenges over the past few years: the powerful pandemic has a fundamental and long-lasting influence on mobility, which is hard to model. Inspired by recent progress in the Large Language Model, we present the Human Mobility Prediction-Large Language Model (HMP-LLM), which utilizes LLMs with understanding and reasoning abilities to do human mobility prediction tasks. Especially to adapt LLMs to intervened time series prediction, we propose the following critical designs: 1) Seasonal Decompose. 2) Two-stage Prompts Designing. We convert the input data into textual prototypes so LLMs can understand it, then we use two-stage prompts to guide the language model in reasoning. Extensive experiments on real-world datasets confirm the superiority of HMP-LLM over existing methods. Besides, due to the zero-shot nature and the expansibility of our two-stage design, HMP-LLM is expected to be used in other public safety incidents.