Research map / Label-efficient and robust learning / Recognition and long-tail
Disentangled Modeling of Preferences and Social Influence for Group 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 DisRec, which disentangles group members' preferences and social influence with separate graph propagation and uses social-importance-based contrastive learning to address preference bias and group data sparsity in group recommendation.
Problem
Group recommendation models entangle member preferences with social influence, causing a preference bias toward majority tastes, and existing SSL augmentations randomly drop members without considering their social importance, while group-item data is sparse.
Approach
Disentangle user-level preference and social-influence embeddings via separate graphs (user-item graph and a social hypergraph with attention over social edges), aggregate members with attention, merge with a group co-occurrence graph via a gate, and generate contrastive pairs by removing the most and least socially influential members.
Key results
- On Mafengwo group recommendation, DisRec reaches HR@10 0.8101 and NDCG@10 0.6312, versus best baseline ConsRec at 0.7829 and 0.5969.
- On Yelp group recommendation, DisRec reaches HR@10 0.7061 and NDCG@10 0.5646, improving NDCG@10 by 9.55% over the best baseline.
- On user recommendation, DisRec improves HR@10 by 12.10% on average over ConsRec across the two datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DisRec instead of the strongest baseline | Group recommendation · Mafengwo | HR@10: 0.7829 -> 0.8101* | Helped |
| Use DisRec instead of the strongest baseline | Group recommendation · Yelp | NDCG@10: 0.5154 -> 0.5646* | Helped |
| Use DisRec instead of the strongest baseline | User recommendation · Mafengwo | HR@10: 0.6346 -> 0.7020* | Helped |
| Use DisRec instead of the strongest baseline | User recommendation · Yelp | NDCG@10: 0.5083 -> 0.5468* | Helped |
| Add social influence embeddings (ablation) | DisRec · Mafengwo group recommendation | HR@10: 0.7698 -> 0.8101* | Helped |
| Add social influence embeddings (ablation) | DisRec · Mafengwo group recommendation | NDCG@10: 0.5916 -> 0.6312* | Helped |
| Add preference embeddings (ablation) | DisRec · Mafengwo group recommendation | HR@10: 0.7888 -> 0.8101* | Helped |
| Add preference embeddings (ablation) | DisRec · Mafengwo group recommendation | NDCG@10: 0.6125 -> 0.6312* | Helped |
| Add social-based SSL (ablation) | DisRec · Mafengwo group recommendation | HR@10: 0.8050 -> 0.8101* | Helped |
| Add social-based SSL (ablation) | DisRec · Mafengwo group recommendation | NDCG@10: 0.6214 -> 0.6312* | Helped |
| Add social influence embeddings (ablation) | DisRec · Yelp group recommendation | HR@10: 0.6625 -> 0.7061* | Helped |
| Add social influence embeddings (ablation) | DisRec · Yelp group recommendation | NDCG@10: 0.5267 -> 0.5646* | Helped |
Limitations
- Hyperparameter tuning ranges are only partially specified; performance sensitive to SSL weight delta.
- Preference bias evaluated only via a synthetic fake-item construction on one dataset.
Experimental setup
Datasets: Mafengwo Yelp
Benchmarks: Mafengwo Yelp
Metrics: HR@5 HR@10 NDCG@5 NDCG@10
Baselines: NCF SIGR HCR GroupIM S$^2$-HHGR SGGCF CubeRec ConsRec
Builds on: AGREE HHGR LightGCN DICE DISGCN
Closest papers by meaning
- Recommending Groups to Users Using User-Group Engagement and Time-Dependent Matrix Factorization (AAAI 2016 · similarity 0.89)
- Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning (AAAI 2025 · similarity 0.87)
- Identify Then Recommend: Towards Unsupervised Group Recommendation (NeurIPS 2024 · similarity 0.86)
- Tail-STEAK: Improve Friend Recommendation for Tail Users via Self-Training Enhanced Knowledge Distillation (AAAI 2024 · similarity 0.86)
- Social Recommendation via Graph-Level Counterfactual Augmentation (AAAI 2025 · similarity 0.85)
- Optimizing Social Network Interventions via Hypergradient-Based Recommender System Design (ICML 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.