Spatial-Aware Case Retrieval Based on Large Language Models: A New Paradigm for Next-Generation POI Recommendation

This article introduces a next-generation POI recommendation method that combines large language models (LLMs) with spatial-aware case retrieval. It breaks through the limitations of traditional sequence models, which ignore spatial relationships and lack interpretability, to achieve more accurate and explainable location recommendations. Through the collaborative work of a spatial encoder, case retrieval module, and LLM reasoner, this method integrates case-based reasoning with semantic understanding capabilities, opening up a new direction for POI recommendation.

Point of Interest (POI) recommendation is a core problem in the field of location services. Traditional methods rely on users' historical check-in sequences and use sequence models such as RNN or Transformer to predict the next location, but they have obvious limitations: they ignore spatial relationships between locations and it is difficult to explain the reasons for recommendations.

The paper proposes a new method of spatial-aware case retrieval, whose core is to combine LLM semantic understanding with case-based reasoning, enabling the model to "draw inferences from one instance". The technical architecture includes three key components:

1. Spatial Encoder: Captures multi-granularity spatial relationships (city → block → POI) through hierarchical spatial representation learning;
2. Case Retrieval Module: Maintains a case library containing user trajectories and next POIs, and retrieves similar cases first when making new recommendations;
3. LLM Reasoner: Understands spatial semantic relationships in cases and generates recommendations by combining structured data.

Case-based reasoning (CBR) has significant advantages in POI recommendation:

• Interpretability: Recommendation results can be traced back to referenced historical cases, allowing users to understand the reasons for recommendations;
• Cold-start friendly: When there is no large amount of user data, recommendations can be made through similar cases;
• Dynamic adaptation: The case library is continuously updated to automatically adapt to new POIs and changes in user behavior.

The paper verifies its effectiveness on multiple real datasets. Compared with traditional sequence models, the recommendation accuracy is significantly improved, especially in scenarios with sparse data and new POI recommendations. This work demonstrates the value of combining LLM semantic understanding with traditional recommendation systems, and the world knowledge of LLMs provides a foundation for spatial reasoning.

The GitHub repository provides complete implementation code (data preprocessing, model training, evaluation scripts), which is valuable for researchers in the following ways:

• Reproducing experimental results;
• Improving and extending based on this work;
• The code structure is clear, making it easy to understand and modify.

Future development directions include: combining multimodal information (POI images, user reviews) to enrich case representation; exploring efficient case retrieval algorithms to handle large-scale case libraries; extending to other sequence recommendation scenarios such as products and music. Conclusion: Spatial-aware case retrieval represents a paradigm shift from "prediction" to "reasoning". With the support of LLMs, recommendation systems become intelligent assistants that can understand context and learn from experience, which is the desired form of next-generation recommendation systems.