PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Burak Tamer, Wolfram Höpken, Zehui Wang

arXiv:2608.16407v1cs.IRcs.LG

TL;DR

Interaction-only POI recommenders struggle to represent businesses with few or no user interactions. LLM-MGCL combines behavioral, LLM-derived semantic, and geographic information in a multi-graph model with contrastive alignment. On the Yelp Multimodal Recommendation Dataset, it improves Recall@20 by 52.0% and NDCG@20 by 64.8% over LightGCN.

  • Problem

    Interaction-only POI recommendation struggles with the cold-start problem, leaving businesses with few or no user interactions with little useful training signal.

    Source: §1 · p. 2

  • Method

    LLM-MGCL extends LightGCN with semantic and geographic item graphs, propagates item representations across all three graphs, and aligns the views through contrastive learning.

    Source: §3 · p. 3, §5 · p. 11

  • Results

    52.0% higher Recall@20 and 64.8% higher NDCG@20 than LightGCN, with LLM-MGCL also performing on par with SGL on the Yelp Multimodal Recommendation Dataset.

    Source: §1 · p. 2, §5 · p. 11

  • Takeaways & Limitations

    The findings suggest externally grounded, LLM-derived item information can help compensate for missing collaborative signal in POI recommendation.

    Source: §5 · p. 11

    A dedicated cold-start evaluation is identified as future work, so the model’s benefit for businesses with few interactions remains to be assessed directly.

    Source: §5 · p. 11

Abstract

from arXiv · show

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

LLM-MGCL addresses cold-start limitations in interaction-based POI recommendation by incorporating LLM-generated business descriptions and geographic proximity into a multi-graph model with contrastive alignment.

1 Introduction

  • Collaborative POI recommenders struggle with cold-start businesses because few or no interactions provide little training signal, while semantic and spatial business information is available.The introduction identifies this item-level information as a potential complement to missing collaborative signals.Source: §1 · p. 2
  • The paper proposes a semantic item-item graph from embeddings of LLM-generated business descriptions and keywords to capture similarity in the Yelp Multimodal Recommendation Dataset.The descriptions are derived from multimodal data, including images and metadata.Source: §1 · p. 2
  • The paper constructs a geographic item-item graph using Haversine distances between business locations to represent spatial proximity relevant to POI recommendation.Source: §1 · p. 2
  • LLM-MGCL integrates user-item interactions with both item-item graphs and is evaluated against collaborative filtering, matrix factorization, and interaction-only graph baselines.The research question is whether textual information and spatial relationships, combined with contrastive alignment, can improve performance over interaction-only approaches.Source: §1 · p. 2

2 Related Work

Related work progresses from interaction-based recommendation toward graph structure, self-supervised signals, and external content; LLM-MGCL combines semantic and geographic item information with LightGCN and contrastive alignment.

GNN-based Recommendation

  • GNN recommenders use bipartite user-item propagation, with LightGCN simplifying graph convolution, but remain dependent on interaction data.Related approaches include NGCF, LightGCN, GC-MC [1], and PinSage.Source: §2 · p. 3

Contrastive Learning

  • SGL adds self-supervision by aligning nodes across interaction-graph views, while SimGCL and NCL use embedding-level perturbations.These contrastive views still derive from the interaction graph's existing information source.Source: §2 · p. 3

Content and Geographic Signals

  • Prior work adds item attributes, multimodal features, LLM-generated semantics, or geographic proximity as complementary recommendation signals.Examples include KGAT, multimodal recommenders, LLM-content work, and geographic signals for POI recommendation.Source: §2 · p. 3

Contrastive Learning

  • Unlike SGL and related methods that contrast perturbed interaction-graph views, LLM-MGCL aligns collaborative representations with semantic and geographic representations.This distinction positions external item knowledge as a contrastive view rather than another perturbation of the interaction graph.Source: §2 · p. 3

3 Methodology

LLM-MGCL extends LightGCN with collaborative, semantic, and geographic graph propagation, using LLM-derived business content and location information to enrich item representations. It aligns representations of the same items across these views through contrastive learning.

3.2 Graph construction

  • The model combines a collaborative user-item graph with semantic and geographic item-item graphs, which share the same item nodes so all three signal types inform item representations.The semantic graph derives from LLM-generated photo summaries; the geographic graph uses Haversine distances between business locations.Source: §3 · p. 4, §3.2 · p. 5

3.1 Dataset and Split

  • Experiments use the Yelp Multimodal Recommendation Dataset and a per-user random split of 80% training, 10% validation, and 10% test interactions.Users with fewer than three interactions are excluded; ChatGPT and DeepSeek-generated summaries and keywords are provided in the processed dataset.Source: §3.1 · p. 4

3.2.1 User-Item Graph

  • The user-item graph assigns interaction edges weights derived from star ratings, mapping 1-star ratings to 0.1 and 5-star ratings to 1.0.A lower clamp keeps low-rated interactions in message passing while emphasizing stronger positive ratings.Source: §3.2.1 · p. 5

3.2.2 SEM-Item-Item Graph

  • The semantic graph encodes concatenated LLM-generated photo summaries and keywords with all-MiniLM-L6-v2, then retains each business’s 10 most similar neighbors.Top-k filtering gives each item, including those with few or no interactions, a consistent semantic signal; edges are bidirectional and similarity-weighted. The source text is attributed to.Source: §3.2.2 · pp. 5–6

3.2.3 GEO-Item-Item Graph

  • The geographic graph retains each business’s 10 nearest neighbors within 50 km and assigns stronger edge weights to closer businesses.Haversine distances determine proximity; weights decrease smoothly with distance and remain between 0 and 1.Source: §3.2.3 · p. 6

3.3 LLM-MGCL Architecture

  • LLM-MGCL propagates shared item embeddings over the user-item graph for three layers and over semantic and geographic item graphs for two layers each, then additively fuses the views.Layer-wise mean pooling includes the initial embedding; the item embedding matrix is shared across paths, and the two auxiliary paths are shallower to avoid diluting signals with loosely related neighbors.Source: §3.3 · pp. 7–8
  • The bidirectional InfoNCE objective aligns each item’s collaborative representation with its semantic and geographic representations, contrasting distinct information sources rather than perturbed copies of one graph.It uses cosine similarity, temperature τ = 0.2, and equal weighting of the two auxiliary views; other items in the batch serve as negatives.Source: §3.3 · p. 8

3.4 LLM-MGCL Training

  • Training combines Bayesian Personalized Ranking, L2 regularization, and contrastive alignment, with λ = 0.05 and γ = 10^-3 scaling the auxiliary terms.For each positive interaction, training samples a user’s positive item uniformly from training interactions and a negative item uniformly from the item set, resampling if it is also positive for that user.Source: §3.4 · pp. 8–9

4 Empirical Results

LLM-MGCL improves recommendation performance over LightGCN and performs comparably overall to SGL, while its ablations identify contrastive alignment as the main driver of gains and show the best results when both auxiliary graphs are combined.

Model comparison

  • Compared with LightGCN, LLM-MGCL raises Recall@20 from 0.0773 to 0.1175 and NDCG@20 from 0.0438 to 0.0722, performing comparably overall to SGL.SGL has higher Recall@10, NDCG@10, and Recall@20, while LLM-MGCL has higher NDCG@20 (0.0722 vs. 0.0712); gaps on the other metrics are small relative to standard deviations.Source: Table 1 · p. 9, §4 · p. 10

Ablation study

  • The ablations isolate contrastive alignment, semantic propagation, and geographic propagation by disabling each component while holding other hyperparameters fixed.The w/o CL variant removes alignment between collaborative and auxiliary views; w/o G_SEM empties the semantic graph, and w/o G_GEO removes geographic propagation.Source: §4 · p. 10
  • Removing contrastive learning lowers Recall@20 from 0.1175 to 0.0906, a 22.9% decrease, making it the primary driver of performance gains.The full model improves over LightGCN by 52.0% in Recall@20 and 64.8% in NDCG@20; the ablations were single runs with seed 42, while the full model was averaged over three seeds.Source: §4 · p. 10, Table 2 · p. 11
  • Removing the semantic or geographic graph yields only minor Recall@20 differences from the full model: 0.1157 and 0.1172, respectively.The authors hypothesize that the graphs encode partially overlapping information because nearby businesses often share semantic characteristics; only the combined model achieves the highest results across all metrics.Source: §4 · p. 10, Table 2 · p. 11

5 Conclusion

LLM-MGCL augments LightGCN with semantic and geographic item graphs and bidirectional contrastive alignment to address item cold-start. On Yelp, it improves over LightGCN and performs on par with SGL; ablations identify cross-view alignment as the primary driver of gains.

5 Conclusion

  • LLM-MGCL extends LightGCN with LLM-summary-and-keyword semantic and Haversine-based geographic item graphs, propagating all three views in parallel and aligning them with bidirectional InfoNCE to address cold-start.Source: §5 · p. 11
  • LLM-MGCL improves LightGCN Recall@20 by 52.0% and NDCG@20 by 64.8%, and performs on par with SGL, the strongest contrastive baseline.The conclusion also reports that it outperforms classical collaborative-filtering, matrix-factorization, and interaction-only graph neural network baselines.Source: §5 · p. 11
  • Removing contrastive alignment reduces Recall@20 by 22.9%, while combining both item-item graphs with alignment achieves the best results across all metrics.Removing either graph alone causes only minor degradation, indicating that semantic edges, spatial proximity, and cross-view alignment provide complementary supervision.Source: §5 · p. 11
  • The authors propose learned view weighting, closer study of each graph’s contribution and overlap, and dedicated evaluation on businesses with few interactions.They suggest externally grounded LLM-derived item information may substitute for missing collaborative signal relied on by models such as SGL.Source: §5 · p. 11

Related papers

Loading 2608.16407v1…