REFINE: A Resource-Efficient LLM-Based Approach for Next Top-K POI Recommendation

Large Language Models (LLMs) have already been proven effective for the next Point-of-Interest (POI) recommendation task due to their ability to capture rich contextual information from historical check-in records. However, limitations such as computational resource requirements and input length constraints make it challenging to include all task-related historical information. Additionally, due to their generation mechanisms and model architectures, standard LLM-based POI recommenders typically produce a single output rather than a ranked Top-K list. Although some methods attempt to generate rankings, they often struggle to produce stable and effective orderings, as the optimization objective of LLMs is not designed for ranking tasks. To address these challenges, we propose a method for leveraging LLMs with designed embedding-based prompts in the next Top-K POI recommendation task. Our method allows LLMs to access all necessary information without the need for the extremely large computational resources required by previous approaches. We evaluate our method on three well-known, real-world open-source datasets, and the results demonstrate that it outperforms other methods on all three datasets.