POI Recommendation with LLM-Augmented Multi-Graph Learning and Contrastive Alignment

Technical Summary: LLM-Augmented Multi-Graph Learning and Contrastive Alignment for POI Recommendation

Problem Statement
Point-of-interest (POI) recommendation systems based on Graph Neural Networks (GNNs), particularly LightGCN, achieve state-of-the-art performance by propagating collaborative signals through user-item interactions. However, these purely collaborative approaches suffer significantly from the cold-start problem. Businesses with few or no user interactions receive little to no training signal, rendering them un-recommendable. While semantic information (e.g., business descriptions, services) and spatial information (e.g., geographic location) are available, they are rarely exploited by standard interaction-based models. This paper addresses the challenge of leveraging these external, item-centric signals to compensate for missing collaborative data.

Methodology: LLM-MGCL
The authors propose LLM-MGCL (LLM-augmented Multi-Graph Contrastive Learning), a multi-graph neural network architecture that extends the LightGCN backbone. The model integrates three distinct graph structures to propagate information in parallel:

1. User-Item Graph ($G_{UI}$): A bipartite graph encoding observed collaborative behavior. Edges are weighted based on star ratings (mapped from 1–5 stars to weights 0.1–1.0) to reflect interaction strength.
2. Semantic Item-Item Graph ($G_{SEM}$): A graph capturing semantic similarity between businesses. It is constructed using LLM-generated photo summaries and keywords from multimodal data. These texts are encoded into embeddings using a Sentence-Transformer model. Edges connect the top-10 most semantically similar businesses, ensuring even cold-start items receive a semantic signal.
3. Geographic Item-Item Graph ($G_{GEO}$): A graph capturing spatial proximity. Distances between business locations are calculated using the Haversine formula. Edges connect the 10 nearest neighbors within a 50 km threshold, with weights inversely proportional to distance.

Architecture and Training

• Multi-Graph Propagation: The model maintains a shared item embedding matrix. It runs three parallel propagation paths:
  • The collaborative path propagates over $G_{UI}$ for 3 layers.
  • The semantic and geographic paths propagate over $G_{SEM}$ and $G_{GEO}$ respectively for 2 layers each (chosen to prevent signal dilution given the fixed $k$-nearest neighbor structure).
• Fusion: The resulting item representations from the three views (Collaborative, Semantic, Geographic) are combined via additive fusion to create a final item embedding.
• Contrastive Alignment: A critical component is the bidirectional InfoNCE contrastive objective. Unlike self-supervised methods that perturb a single graph (e.g., SGL), LLM-MGCL aligns the collaborative view with the two auxiliary views (semantic and geographic). This forces the model to learn consistent representations for the same item across behavioral, semantic, and spatial contexts.
• Loss Function: The total loss combines the Bayesian Personalized Ranking (BPR) loss for ranking, the contrastive loss for view alignment, and an L2 regularization term.

Key Contributions

1. Construction of LLM-Derived Semantic Graphs: The paper introduces a method to build item-item graphs using LLM-generated summaries and keywords, capturing semantic similarity independent of user interactions.
2. Integration of Geographic Proximity: The construction of a geographic item-item graph based on Haversine distances, explicitly modeling the spatial constraints relevant to POI recommendations.
3. Multi-Graph Contrastive Framework: The design of a model that integrates these auxiliary graphs with collaborative filtering and aligns them via contrastive learning, specifically targeting the cold-start problem.

Experimental Results
Experiments were conducted on the Yelp Multimodal Recommendation Dataset. The model was compared against classical collaborative filtering (ItemKNN, UserKNN), matrix factorization (MF-BPR, NeuMF), and various GNN baselines (LightGCN, NGCF, GC-MC, SGL).

• Performance Gains: LLM-MGCL significantly outperformed the LightGCN backbone, improving Recall@20 by 52.0% and NDCG@20 by 64.8%.
• Comparison with SGL: The model performed on par with SGL (Self-supervised Graph Learning), the strongest contrastive baseline. While SGL achieved slightly higher Recall@10 and NDCG@10, LLM-MGCL achieved a higher NDCG@20. Crucially, unlike SGL, LLM-MGCL utilizes external signals to address items with no interactions.
• Ablation Study:
  • Contrastive Learning (CL): Removing the contrastive alignment caused a substantial drop in performance (Recall@20 decreased by ~22.9%), identifying it as the primary driver of gains.
  • Auxiliary Graphs: Removing either the semantic or geographic graph individually resulted in only minor performance degradation, suggesting these views encode partially overlapping information (e.g., businesses in the same area often share semantic traits). However, the best performance was achieved only when both graphs were combined with contrastive learning, confirming their complementary nature.

Significance and Claims
The paper claims that externally grounded, LLM-derived item knowledge can effectively compensate for missing collaborative signals. By aligning behavioral data with semantic and spatial representations, the model mitigates the item cold-start problem inherent in purely interaction-based methods. The authors suggest that this approach offers a viable path for improving POI recommendations where interaction data is sparse, without relying solely on the self-supervised perturbations of existing interaction graphs. Future work directions identified by the authors include exploring learned weighting mechanisms for view fusion, deeper analysis of information overlap between graphs, and dedicated evaluations for cold-start scenarios.