Large Language Models for Next Point-of-Interest Recommendation

Abstract

The next Point of Interest (POI) recommendation task is to predict users'\nimmediate next POI visit given their historical data. Location-Based Social\nNetwork (LBSN) data, which is often used for the next POI recommendation task,\ncomes with challenges. One frequently disregarded challenge is how to\neffectively use the abundant contextual information present in LBSN data.\nPrevious methods are limited by their numerical nature and fail to address this\nchallenge. In this paper, we propose a framework that uses pretrained Large\nLanguage Models (LLMs) to tackle this challenge. Our framework allows us to\npreserve heterogeneous LBSN data in its original format, hence avoiding the\nloss of contextual information. Furthermore, our framework is capable of\ncomprehending the inherent meaning of contextual information due to the\ninclusion of commonsense knowledge. In experiments, we test our framework on\nthree real-world LBSN datasets. Our results show that the proposed framework\noutperforms the state-of-the-art models in all three datasets. Our analysis\ndemonstrates the effectiveness of the proposed framework in using contextual\ninformation as well as alleviating the commonly encountered cold-start and\nshort trajectory problems.\n