Research map / Graph learning / Knowledge and temporal graphs
LGMRec: Local and Global Graph Learning for Multimodal Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LGMRec, a multimodal recommender that decouples collaborative and modality graph embeddings on local user-item graphs and adds a hypergraph module with learnable implicit attribute hyperedges to capture global user interests.
Problem
Multimodal recommendation methods share user ID embeddings between collaborative and modality signals (coupling), and model only local interaction-graph interests, which suffers under sparse interactions (locality).
Approach
Independently learn collaborative and modality-related local embeddings (user modal features initialized by averaging their interacted items' modal features), and learn global user/item embeddings via a hypergraph whose hyperedges are learnable implicit attributes, with Gumbel-Softmax augmentation and cross-modal hypergraph contrastive learning.
Key results
- On Baby, LGMRec achieves R@20 0.1002 and N@20 0.0440, versus the strongest baseline MICRO at 0.0905 and 0.0406.
- On Sports, LGMRec reaches R@10 0.0720 and R@20 0.1068, improving on the best baseline by 6.51% and 4.09% in R@10 and R@20.
- On Clothing, LGMRec improves R@10 by 11.90% over the best baseline, and LGMRec gains 19.95% over the best baseline on the sparsest user group of Baby.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add multimodal contents (ablation) | LGMRec · Baby, Sports, Clothing | Recall@20: 0.0732 -> 0.1002; 0.0829 -> 0.1068; 0.0514 -> 0.0828 | Helped |
| Add local graph embedding (ablation) | LGMRec · Baby, Sports, Clothing | Recall@20: 0.0806 -> 0.1002; 0.0851 -> 0.1068; 0.0741 -> 0.0828 | Helped |
| Add collaborative graph embedding alongside modality graph embedding (ablation) | LGMRec · Baby, Sports, Clothing | Recall@20: 0.0947 -> 0.1002; 0.0997 -> 0.1068; 0.0807 -> 0.0828 | Helped |
| Add modality graph embedding alongside collaborative graph embedding (ablation) | LGMRec · Baby, Sports, Clothing | Recall@20: 0.0929 -> 0.1002; 0.0988 -> 0.1068; 0.0804 -> 0.0828 | Helped |
| Add global hypergraph embedding (ablation) | LGMRec · Baby, Sports, Clothing | Recall@20: 0.0972 -> 0.1002; 0.1032 -> 0.1068; 0.0803 -> 0.0828 | Helped |
| Add hypergraph contrastive learning (ablation) | LGMRec · Baby, Sports, Clothing | Recall@20: 0.0992 -> 0.1002; 0.1051 -> 0.1068; 0.0812 -> 0.0828 | Helped |
| Keep user ID embeddings decoupled between CGE and MGE instead of sharing them (ablation) | LGMRec · Baby, Sports, Clothing | Recall@20: 0.0869 -> 0.1002; 0.0895 -> 0.1068; 0.0713 -> 0.0828 | Helped |
Surprising findings
- About 94.26% of users in the Baby dataset show more than 50% of embedding dimensions with opposite gradient directions between collaborative and multimodal modules early in training, indicating coupling of signals.
Limitations
- Modality fusion is limited to visual and textual modalities; the authors note that better modeling of differences and commonalities among modalities is future work.
- Hyperparameters (layers, hyperedge count, alpha, lambda) are tuned by grid search, with different optimal settings per dataset.
- Reliance on pre-extracted modal features (4096-d visual, 384-d textual) from prior work.
Experimental setup
Datasets: Amazon Baby Amazon Sports and Outdoors Amazon Clothing Shoes and Jewelry
Benchmarks: Amazon Baby Amazon Sports and Outdoors Amazon Clothing Shoes and Jewelry
Metrics: Recall@10 Recall@20 NDCG@10 NDCG@20
Baselines: BPR LightGCN SGL NCL HCCF SHT VBPR MMGCN GRCN LATTICE MMGCL MICRO
Builds on: LightGCN SGL HCCF BPR MMRec
Closest papers by meaning
- Modality-Independent Graph Neural Networks with Global Transformers for Multimodal Recommendation (AAAI 2025 · similarity 0.90)
- DOGE: LLMs-Enhanced Hyper-Knowledge Graph Recommender for Multimodal Recommendation (AAAI 2025 · similarity 0.90)
- Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 2025 · similarity 0.90)
- MENTOR: Multi-level Self-supervised Learning for Multimodal Recommendation (AAAI 2025 · similarity 0.90)
- Curriculum Conditioned Diffusion for Multimodal Recommendation (AAAI 2025 · similarity 0.89)
- STAIR: Manipulating Collaborative and Multimodal Information for E-Commerce Recommendation (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.