BLCSRec: Bridging Language and Collaborative Semantics for Next POI Recommendation

Yibin Liu, Shijian Li, Jianyu Zhang

Keywords: Recommender System, Large language model, Generative Recommenders

TL;DR: Bridging Language and Collaborative Semantics for Next POI Recommendation

Abstract: Next Point-of-Interest (POI) recommendation seeks to predict a user’s future location based on their past mobility patterns, a task essential for personalized location-based services. To enhance semantic modeling in this context, recent studies have applied large language models (LLMs), where early approaches relied on randomly assigned POI identifiers with limited representational capacity. More recent work has introduced semantic identifiers (SIDs) to better capture spatial and contextual correlations, leading to improved prediction accuracy. However, these approaches face several critical challenges: a semantic gap between LLM's language semantics and POI-specific collaborative semantics, and the generation of invalid SIDs that do not correspond to real POIs. To tackle these challenges, we propose a novel framework called BLCSRec for bridging language and collaborative semantics in next POI recommendation. Specifically, we introduce an LLM-based POI profile generation method that summarizes user trajectories and integrates POI attributes with visitor information, and further employ an RQ-VAE to encode addresses, category, and these enriched textual profiles into semantic identifiers that capture both static attributes and collaborative context. To alleviate the gap between language and collaborative semantic, we incorporate explicit alignment tasks that map SIDs to/from textual descriptions and implicit alignment tasks that predict next POIs in asymmetric semantic formats. Furthermore, we employ GRPO reinforcement learning with a hierarchical reward structure to suppress invalid SID generation and enhance accuracy. Extensive experiments on three public datasets (NYC, TKY, and CA) demonstrate that our method consistently outperforms strong LLM-based baselines, achieving improvements of 7.3\% in Acc@1 on NYC and 3.5\% on CA, along with substantial reductions in invalid SID generation.

Primary Area: generative models

Submission Number: 9417