Collaborative Semantics-Assisted Large Language Models for Next POI Recommendation

Abstract

Next point-of-interest (POI) recommendation aims to forecast users’ next POI visit based on their historical movement information. Existing methods typically explore latent transition patterns within complex human activity trajectories by sequential or graph-based paradigms. However, they essentially oversimplify transition patterns by treating users and POIs as nodes, learning uniform embeddings for each ID. This results in a vast of important irregular textual semantic information being overlooked. To this end, we propose a novel large language model-based Collaborative Semantics-Assisted model called CSA-Rec for next POI recommendation. Specifically, our model retains the irregular check-in information in its original format to prevent the loss of important contextual details. These textual information are subsequently mapped to new, meaningful POI IDs through a meticulously crafted dual-channel vector quantization module. Then, we devise two fine-tuning tasks to reinforce the integration of collaborative trajectory signals and POI semantic information in the large language models. Extensive experiments on several real-world datasets validate the effectiveness of our CSA-Rec.