LLM-enhanced semantic–collaborative fusion for next POI recommendation
Abstract Next point-of-interest (POI) recommendation is an important task in location-based social networks and intelligent mobility services. Existing sequential and graph-based models mainly rely on discrete IDs and structured propagation, limiting their ability to capture activity semantics and immediate mobility intentions behind check-in behaviors. Although large language models (LLMs) have strong semantic understanding capabilities, they may suffer from task mismatch and hallucination issues when directly applied to recommendation scenarios. To address these limitations, LSCF-POI is proposed as an LLM-enhanced semantic–collaborative fusion framework for next POI recommendation. Specifically, a trajectory prompting module is designed to convert structured check-in records into natural language narratives containing visit order, spatiotemporal constraints, and POI category transitions, enabling an LLM to infer users’ fine-grained dynamic mobility intentions through semantic reasoning. Meanwhile, LightGCN is adopted to extract macro-level collaborative preferences from the global user–POI interaction graph. Finally, an adaptive attention mechanism is introduced to align unstructured semantic reasoning with structured collaborative topology in a unified latent space. Extensive experiments on real-world datasets demonstrate that LSCF-POI consistently outperforms competitive baselines. These results demonstrate the complementary value of integrating micro-level semantic intentions with macro-level collaborative awareness for next POI recommendation.
Authors
- Hongyu An (ORCID: https://orcid.org/0009-0008-5551-1242)
- Qi Jiang
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1038/s41598-026-79376-3
- Primary Topic
- Recommender Systems and Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00