Research map / Label-efficient and robust learning / Recognition and long-tail
Identify Then Recommend: Towards Unsupervised 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 ITR, an unsupervised group recommendation framework that identifies user groups without a preset group count via adaptive density estimation and a merge-and-split strategy, then trains with pull-and-repulsion and pseudo-group recommendation pre-text tasks.
Problem
Group recommendation relies on a pre-defined, fixed number of user groups and on costly user-group and group-item annotations, which fail under dynamic industrial group distributions.
Approach
Identify user groups unsupervised by estimating adaptive density of user embeddings (high-density points become group centers), refine groups with a heuristic explore-merge-split procedure, and train with self-supervised pull-and-repulsion and pseudo group-item labels derived from estimated user-group assignments.
Key results
- Group recommendation on Mafengwo: ITR NDCG@5 0.8761 vs. best baseline CoHeat 0.7924, a 0.0837 absolute gain.
- On CAMRa2011 group recommendation, ITR NDCG@5 0.6653, a 22.95% relative improvement over the runner-up as reported in the paper.
- A/B test on a live industrial recommender shows ITR improving the baseline by about 2.45% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ITR instead of CoHeat | Group recommendation · Mafengwo | NDCG@5: 0.7924 -> 0.8761 | Helped |
| Use ITR instead of ConsRec | Group recommendation · CAMRa2011 | NDCG@5: 0.4358 -> 0.6653 | Helped |
| Use ITR instead of MeLU | User recommendation · Mafengwo | NDCG@5: 0.7095 -> 0.7569 | Helped |
| Use ITR instead of ConsRec | User recommendation · CAMRa2011 | NDCG@5: 0.4568 -> 0.6790 | Helped |
| Add GIM and pull-and-repulsion task (ablation) | Base group recommender · Mafengwo | NDCG@5: 0.7692 -> 0.8152 | Helped |
| Add GIM and pseudo group recommendation task (ablation) | Base group recommender · Mafengwo | NDCG@5: 0.7692 -> 0.7934 | Helped |
| Add GIM and pull-and-repulsion task (ablation) | Base group recommender · CAMRa2011 | NDCG@5: 0.4358 -> 0.6345 | Helped |
| Add GIM and pseudo group recommendation task (ablation) | Base group recommender · CAMRa2011 | NDCG@5: 0.4358 -> 0.5559 | Helped |
| Combine PAR and PGR with GIM (ablation) | Base group recommender · Mafengwo | NDCG@5: 0.7692 -> 0.8761 | Helped |
| Combine PAR and PGR with GIM (ablation) | Base group recommender · CAMRa2011 | NDCG@5: 0.4358 -> 0.6653 | Helped |
| Add GIM and pull-and-repulsion task (ablation) | Base user recommender · Mafengwo | NDCG@5: 0.6884 -> 0.7125 | Helped |
| Add GIM and pseudo group recommendation task (ablation) | Base user recommender · Mafengwo | NDCG@5: 0.6884 -> 0.7235 | Helped |
Surprising findings
- Removing group annotations causes ConsRec's group recommendation HR@5 on Mafengwo to fall from 0.8844 to 0.0563, while user recommendation is barely affected.
- ITR does not beat GroupIM on CAMRa2011 HR@10, which the authors attribute to GroupIM's ranking ability.
Limitations
- Still depends on user-item interactions for both user and group recommendation.
- Evaluated only on two public datasets plus a proprietary industrial deployment without public release.
- Several design choices (e.g., heuristic merge-and-split, Gaussian explore step) are hand-crafted with limited theoretical justification.
Experimental setup
Datasets: Mafengwo CAMRa2011
Benchmarks: Mafengwo CAMRa2011
Metrics: HR@5 HR@10 NDCG@5 NDCG@10
Baselines: Pop NCF AGREE MAML GroupIM HyperGroup HCR HHGR MeLU CubeRec ConsRec CoHeat
Builds on: ConsRec DBSCAN DeepDPM
Closest papers by meaning
- RecFlow: An Industrial Full Flow Recommendation Dataset (ICLR 2025 · similarity 0.87)
- Disentangled Modeling of Preferences and Social Influence for Group Recommendation (AAAI 2025 · similarity 0.86)
- Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework (AAAI 2024 · similarity 0.86)
- Multifaceted User Modeling in Recommendation: A Federated Foundation Models Approach (AAAI 2025 · similarity 0.84)
- Tail-STEAK: Improve Friend Recommendation for Tail Users via Self-Training Enhanced Knowledge Distillation (AAAI 2024 · similarity 0.84)
- Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.84)
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.