Source-linked AI summary
POI Recommendation with LLM-Augmented Multi-Graph Learning and Contrastive Alignment
Burak Tamer, Wolfram Höpken, Zehui Wang
TL;DR
POI recommenders relying on user-item interactions struggle to represent businesses with few or no interactions. LLM-MGCL adds semantic and geographic item graphs to LightGCN and aligns their representations contrastively, improving Recall@20 by 52.0% and NDCG@20 by 64.8% over LightGCN.
Problem
Interaction-only POI recommenders struggle with cold-start businesses because items with few or no user interactions receive little useful training signal.
Method
LLM-MGCL extends LightGCN with LLM-derived semantic and geographic item graphs, propagates representations across all views, and aligns them using contrastive learning.
Results
LLM-MGCL improves Recall@20 by 52.0% and NDCG@20 by 64.8% over LightGCN while performing on par with SGL.
Takeaways & Limitations
Externally grounded, LLM-derived item information can help compensate for missing collaborative signal and mitigate cold-start problems in POI recommendation.
Takeaways & Limitations
The model’s cold-start benefit is not directly assessed because the paper identifies dedicated evaluation for low-interaction businesses as future work.
Abstract
from arXiv · showhide
Point-of-interest (POI) recommendation models based on graph neural networks achieve strong performance by propagating collaborative signals over user-item interactions, yet they struggle with the cold-start problem, where items with few or no interactions are not represented. In this paper, we propose LLM-augmented Multi-Graph Contrastive Learning (LLM-MGCL), a multi-graph neural network that uses semantic and spatial information about items to extend the LightGCN backbone with two auxiliary item-item graphs: a semantic graph constructed from sentence embeddings of LLM-generated photo summaries and keywords, and a geographic graph derived from Haversine distances between business locations. Item embeddings are propagated over all three graphs in parallel, fused additively, and aligned across views through a bidirectional InfoNCE contrastive objective that connects behavioral, semantic, and spatial representations of the same items. Experiments on the Yelp Multimodal Recommendation Dataset show that LLM-MGCL outperforms classical collaborative filtering, matrix factorization, and interaction-only graph neural network baselines. It improves Recall@20 by 52.0% and NDCG@20 by 64.8% over LightGCN while performing on par with the strongest contrastive baseline, Self-supervised Graph Learning (SGL), which is also affected by the cold-start problem. An ablation study reveals that the cross-view contrastive alignment (CA) is the primary driver of these gains, with the best performance achieved when all three graphs are combined. Our results suggest that externally grounded, LLM-derived item knowledge can effectively compensate for missing collaborative signal and mitigate the item cold-start problem in POI recommendation.
1 Introduction
The paper addresses POI recommendation’s limitations by augmenting LightGCN with LLM-derived semantic information, geographic proximity, and contrastive multi-graph learning. It introduces semantic and geographic item-item graphs alongside user-item interactions to improve recommendation performance.
- LightGCN propagates collaborative signals across user-item interactions and has emerged as a strong backbone for POI recommendation.
- LLM-generated business descriptions from images and metadata, together with geographic proximity, are proposed to enhance POI recommendation.
- LLM-MGCL constructs semantic and geographic item-item graphs, capturing textual similarity and Haversine-based spatial proximity between businesses.
2 Related Work
Related work progresses from interaction-only collaborative filtering toward models incorporating auxiliary graphs, self-supervision, and external content. LLM-MGCL builds on LightGCN by adding LLM-generated semantics and geographic proximity, aligned contrastively rather than through randomly perturbed views.
- Related work: Recommender systems increasingly incorporate auxiliary graph structure, self-supervised signals, and external content to address limitations of interaction-based collaborative filtering.These developments motivate LLM-MGCL’s design choices.
- GNN-based Recommendation: Graph neural networks have become a dominant collaborative-filtering paradigm by modeling user-item interactions as a bipartite graph.The passage identifies NGCF as introducing standard graph convolution, though the provided text is truncated.
- Novelty: LLM-MGCL extends LightGCN with LLM-generated semantic content and geographic proximity, aligning these views through contrastive learning.Its contrastive objective differs from SGL and successors by contrasting collaborative representations with semantic and geographic representations instead of randomly perturbed views.
3 Methodology
LLM-MGCL extends LightGCN with parallel behavioral, semantic, and geographic graph propagation, using LLM-generated content and business locations to enrich item representations. It aligns these views through symmetric contrastive learning so collaborative behavior is connected with externally grounded semantic and spatial information.
- Model architecture: LLM-MGCL extends LightGCN with three parallel propagation paths over user-item, semantic item-item, and geographic item-item graphs.The three graphs share the same item node set, allowing behavioral, semantic, and geographic signals to influence a common item representation.
- Graph construction: The user-item graph encodes collaborative behavior, while auxiliary item-item graphs represent LLM-derived semantic similarity and geographic proximity.The semantic graph uses LLM-generated photo summaries and keywords; the geographic graph is derived from Haversine distances between business locations.
- Semantic graph: k = 10 semantic neighbors are retained per business, giving every item, including those with few or no interactions, a consistent semantic propagation signal.Semantic representations concatenate LLM-generated photo summaries and keywords and encode them with all-MiniLM-L6-v2 into 384-dimensional L2-normalized embeddings.
- Geographic graph: Geographic edges retain businesses within 50 km and the 10 nearest neighbors per business to form balanced spatial neighborhoods.The graph represents locational relationships relevant to POI recommendation and uses distance-based edge weighting followed by symmetric LightGCN normalization.
- Contrastive alignment: Contrastive alignment links behavioral co-occurrence with semantic or geographic knowledge rather than contrasting randomly perturbed interaction-graph views.The loss is computed symmetrically in both directions, averages both views equally, and uses τ = 0.2 to distinguish similar from dissimilar pairs.
4 Empirical Results
LLM-MGCL outperforms collaborative, factorization, and interaction-only GNN baselines on Yelp, while ablations identify contrastive learning as the main performance driver and show complementary value from semantic and geographic item graphs.
- Model comparison: LLM-MGCL outperforms classical, matrix factorization, and collaborative-filtering-only GNN methods, including LightGCN, NGCF, and GC-MC.The comparison covers ItemKNN, UserKNN, MF-BPR, NeuMF, LightGCN, NGCF, GC-MC, and SGL.
- Model comparison: 52.0%: Recall@20 improves from 0.0773 with LightGCN to 0.1175 with LLM-MGCL.LightGCN is the direct backbone of LLM-MGCL.
- Model comparison: 64.8%: NDCG@20 improves from 0.0438 with LightGCN to 0.0722 with LLM-MGCL.The passage describes this as a gain of roughly 65%.
- Ablation study: Removing contrastive learning reduces Recall@20 from 0.1175 to 0.0906, a relative decrease of 22.9%, while both item-item graphs remain.This identifies contrastive learning as the primary driver of performance gains.
- Ablation study: Removing the semantic graph yields Recall@20 of 0.1157, whereas removing the geographic graph yields 0.1172, producing only minor performance differences.The ablation results suggest both auxiliary graphs contribute less than contrastive alignment to the observed gains.
5 Conclusion
The paper proposes LLM-MGCL, which augments LightGCN with semantic and geographic item-item graphs and aligns collaborative, semantic, and geographic views. On Yelp, it outperforms traditional and interaction-only baselines, improving LightGCN results while matching SGL.
- Method: LLM-MGCL extends LightGCN with semantic and geographic item-item graphs built from LLM-generated item information and Haversine business-location distances.Item embeddings are propagated over all three graphs in parallel and aligned across views.
- Results: 52.0% and 64.8%: LLM-MGCL improves LightGCN on Recall@20 and NDCG@20, respectively, on the Yelp Multimodal Recommendation Dataset.It also substantially outperforms classical collaborative filtering, matrix factorization, and interaction-only graph neural network baselines.
- Implications: LLM-MGCL performs on par with SGL, the strongest contrastive baseline, while externally grounded LLM-derived item information offers a path to mitigating POI recommendation cold-start problems.The approach is presented as a substitute for missing collaborative signal on which models such as SGL rely.