Research map / Graph learning / GNN architectures
Mind Individual Information! Principal Graph Learning for Multimedia 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 Principal Graph Learning (PGL), which extracts a principal subgraph from the user-item graph (global-aware SVD truncation or local-aware exposure-weighted sampling) for GNN message passing plus a feature-masking InfoNCE auxiliary task, to better capture individual information in multimedia recommendation.
Problem
GNN-based multimedia recommenders capture co-occurrence (low-frequency, global) patterns but underutilize local structural features carrying individual information, leading to over-similar user preference representations.
Approach
Message passing is performed on a principal subgraph extracted from the user-item graph during training (via global-aware truncated SVD reconstruction or local-aware exposure-weighted sampling), while inference uses the full graph; a feature-discriminating self-supervised task further improves representation distinguishability.
Key results
- PGL with local-aware extraction reaches Recall@20 of 0.1022 on Baby, 0.1174 on Sports and 0.1034 on Clothing, outperforming the strongest baselines (LGMRec, MGCN, FREEDOM) on all three datasets.
- The paper reports an average improvement of about 9% over existing methods.
- Ablation shows removing principal graph learning significantly reduces performance; removing the SSL task also lowers performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add principal graph learning in place of complete-graph learning (ablation) | PGL · Amazon recommendation datasets | Recall@20 | Helped |
| Add the self-supervised auxiliary task (ablation) | PGL · Amazon recommendation datasets | Recall@20 | Helped |
| Use local-aware rather than global-aware principal subgraph extraction | PGL · Baby | Recall@20: 0.1040 -> 0.1022 | Mixed |
| Use local-aware rather than global-aware principal subgraph extraction | PGL · Sports | Recall@20: 0.1144 -> 0.1174 | Helped |
| Use local-aware rather than global-aware principal subgraph extraction | PGL · Clothing | Recall@20: 0.1014 -> 0.1034 | Helped |
Surprising findings
- Local-aware extraction (exposure-weighted sampling) outperforms the global-aware SVD-based extraction, which the authors attribute to partially alleviating exposure bias.
Limitations
- Gains reported on only three Amazon categories with pre-extracted visual/text features.
- Optimal hyperparameters (masking ratio, SSL weight, sampling ratio p) vary across datasets and require tuning.
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@K NDCG@K Uniformity
Baselines: MF-BPR LightGCN DirectAU LayerGCN VBPR MMGCN DualGNN LATTICE SLMRec MICRO BM3 MMSSL
Builds on: LATTICE LightGCN FREEDOM MGCN DirectAU
Closest papers by meaning
- Modality-Independent Graph Neural Networks with Global Transformers for Multimodal Recommendation (AAAI 2025 · similarity 0.86)
- Spectral-Based Graph Neural Networks for Complementary Item Recommendation (AAAI 2024 · similarity 0.85)
- LGMRec: Local and Global Graph Learning for Multimodal Recommendation (AAAI 2024 · similarity 0.85)
- ContextGNN: Beyond Two-Tower Recommendation Systems (ICLR 2025 · similarity 0.85)
- Hgformer: Hyperbolic Graph Transformer for Collaborative Filtering (ICML 2025 · similarity 0.85)
- Semantic Enhanced Heterogeneous Hypergraph Network for Collaborative Filtering (AAAI 2025 · similarity 0.85)
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.