Adapting Large Language Model for Spatio-Temporal Understanding in Next Point-of-Interest Prediction

Abstract

The widespread deployment of Large Language Models (LLMs) across various sectors has underscored their versatility beyond conventional natural language processing applications. Although LLMs are adept at analyzing time series and geospatial data, their capacity to process human spatio-temporal activity data is not yet fully explored. To bridge this research gap, we introduce "LLM-Next," a model engineered to understand spatio-temporal data for predicting a user’s next Point-of-Interest (POI). We developed a specific data processing approach optimized for both temporal and spatial data and fine-tuned the LLM to improve its contextual awareness of human mobility patterns. Comparative experiments on three distinct datasets of human mobility in the real world demonstrate that LLM-Next outperforms existing baseline models in accuracy, thus redefining benchmarks in the domain.