Research map / Label-efficient and robust learning / Recognition and long-tail
Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning
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 RGCML, a graph-based social recommender that denoises social ties with HK opinion dynamics, fuses multi-view auxiliary signals (social, implicit social, global intents) via personalized low-rank transfer matrices, and adds global-local InfoNCE contrastive learning.
Problem
Social recommendation suffers from noisy social connections and coarse integration of auxiliary information (social and intent signals), which biases user preferences under data sparsity.
Approach
Uses Hegselmann-Krause opinion dynamics (cosine-similarity thresholding of opinion embeddings) to filter unreliable social ties, combines social and global-intent embeddings through a personalized fusion module that generates user-specific transfer matrices via low-rank factorization, and aligns final user embeddings with local (preference) and global (intent) views using contrastive losses.
Key results
- On Yelp, RGCML reaches Recall@5 0.05345 and NDCG@5 0.0596, versus 0.05084 and 0.05834 for the best baseline SGL.
- On Ciao, RGCML improves Recall@5 by up to 3.5% and NDCG@5 by 2.53% over baselines, with the largest gains on the sparser dataset.
- Removing both contrastive modules (w/o-both) drops NDCG@5 by 11.6% on Ciao, 17.3% on Yelp and 11.9% on Douban.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RGCML instead of the strongest baseline | Social recommendation · Yelp | Recall@5: 0.05084 -> 0.05345 | Helped |
| Use RGCML instead of the strongest baseline | Social recommendation · Yelp | NDCG@5: 0.05834 -> 0.0596 | Helped |
| Use RGCML instead of the strongest baseline | Social recommendation · Ciao | Recall@5: 0.02915 -> 0.03017 | Helped |
| Use RGCML instead of the strongest baseline | Social recommendation · Ciao | NDCG@5: 0.03715 -> 0.03809 | Helped |
| Use RGCML instead of the strongest baseline | Social recommendation · Douban | Recall@20: 0.17696 -> 0.18388 | Helped |
| Add implicit social relations (ablation) | RGCML · Yelp, Ciao, and Douban | Recall@5: Yelp 0.05228 -> 0.05345; Ciao 0.02969 -> 0.03017; Douban 0.08484 -> 0.08503 | Mixed |
| Add explicit social information (ablation) | RGCML · Yelp, Ciao, and Douban | Recall@5: Yelp 0.05255 -> 0.05345; Ciao 0.02906 -> 0.03017; Douban 0.08438 -> 0.08503 | Helped |
| Use the designed information fusion module instead of an MLP (ablation) | RGCML · Yelp, Ciao, and Douban | Recall@5: Yelp 0.05175 -> 0.05345; Ciao 0.02852 -> 0.03017; Douban 0.07564 -> 0.08503 | Helped |
| Add local contrastive learning (ablation) | RGCML · Yelp, Ciao, and Douban | Recall@5: Yelp 0.05155 -> 0.05345; Ciao 0.02802 -> 0.03017 | Helped |
| Add global contrastive learning (ablation) | RGCML · Yelp, Ciao, and Douban | NDCG@5: Yelp 0.0592 -> 0.0596; Ciao 0.03719 -> 0.03809; Douban 0.1478 -> 0.1492 | Mixed |
| Add both contrastive learning components (ablation) | RGCML · Yelp, Ciao, and Douban | NDCG@5: Yelp 0.05083 -> 0.0596; Ciao 0.03413 -> 0.03809; Douban 0.13338 -> 0.1492 | Helped |
Surprising findings
- Adding social connections does not always help: on Yelp, which has many social ties, MHCN and SEPT underperform SGL.
- Social information has little measurable effect on Douban, where removing social relations barely changes performance.
Limitations
- Evaluated only on three datasets with a single 7:1:2 split and top-N metrics.
- Hyperparameters such as K, alpha, lambdas and temperatures are tuned by grid search; sensitivity depends on dataset.
Experimental setup
Datasets: Ciao Yelp Douban
Benchmarks: Ciao Yelp Douban
Metrics: Recall@5 Recall@20 NDCG@5 NDCG@20
Baselines: LightGCN DiffNet ESRF SGL SEPT MHCN NCL DCCF
Builds on: LightGCN SGL NCL DCCF Hegselmann-Krause opinion dynamics
Closest papers by meaning
- DivGCL: A Graph Contrastive Learning Model for Diverse Recommendation (AAAI 2025 · similarity 0.89)
- Disentangled Modeling of Preferences and Social Influence for Group Recommendation (AAAI 2025 · similarity 0.87)
- Disentangled Contrastive Bundle Recommendation with Conditional Diffusion (AAAI 2025 · similarity 0.87)
- DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation (AAAI 2025 · similarity 0.87)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.87)
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.87)
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.