Research map / Graph learning / Graph self-supervised learning
DivGCL: A Graph Contrastive Learning Model for Diverse 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 DivGCL, a LightGCN-based graph contrastive recommender that adds a DPP likelihood loss for diversity and a Gaussian-noise-augmented CL loss to balance accuracy and diversity.
Problem
GCL-based recommenders favor popular items and sacrifice diversity; existing diversification methods are mostly post-processing and disconnected from end-to-end embedding learning, so accuracy and diversity are hard to balance.
Approach
Combines a determinant point process (DPP) likelihood loss over positive and negative item sets, which models quality and pairwise similarity, with an InfoNCE-style contrastive loss where the augmented view is generated by Gaussian noise and contrasted against the main view.
Key results
- On ML-1M, DivGCL improves T@20 by 23.47% over the sub-optimal SimGCL (10.26 vs 8.31).
- DivGCL improves Recall@20 by 18.49% and NDCG@20 by 24.39% on ML-1M over the best baseline.
- On Beauty, DivGCL raises T@20 by 2.78% over SGL, but is slightly inferior to SimGCL on ILAD@20 and ILMD@20.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DivGCL instead of SimGCL | LightGCN recommender · ML-1M | T@20: 8.31 -> 10.26; 23.47% | Helped |
| Use DivGCL instead of SGL | LightGCN recommender · Beauty | T@20: 16.11 -> 16.57; 2.78% | Helped |
| Use DivGCL instead of SimGCL | LightGCN recommender · AMiner | T@20: 17.99 -> 19.67; 9.34% | Helped |
| Use DivGCL instead of SGL | LightGCN recommender · Yelp2018 | T@20: 9.83 -> 10.73; 9.16% | Helped |
| Add the DPP likelihood-based loss (ablation) | DivGCL · ML-1M | T@20: 10.05 -> 10.26 | Helped |
| Add the DPP likelihood-based loss (ablation) | DivGCL · Beauty | T@20: 16.27 -> 16.57 | Helped |
| Add the noise-based contrastive learning objective (ablation) | DivGCL · ML-1M | T@20: 7.43 -> 10.26 | Helped |
| Add the noise-based contrastive learning objective (ablation) | DivGCL · Beauty | T@20: 10.12 -> 16.57 | Helped |
| Vary embedding size d (ablation) | DivGCL · ML-1M | T@20 | Mixed |
| Increase perturbation strength η (ablation) | DivGCL · ML-1M | T@20 | Mixed |
| Increase DPP loss weight α (ablation) | DivGCL · ML-1M | T@20 | Mixed |
| Vary graph convolution layers (ablation) | DivGCL · ML-1M | T@20 | Mixed |
Surprising findings
- Diversity metrics (ILAD/ILMD) are not optimal for DivGCL on any of the four datasets; the authors accept this as a trade-off for accuracy gains.
- Removing the CL component lowers both accuracy and diversity, suggesting the DPP loss alone does not suffice.
Limitations
- Diversity metrics (ILAD, ILMD) are not best on any dataset, so diversity is traded for accuracy.
- Hyperparameters (noise strength, alpha, layers) are sensitive; excessive noise or alpha hurts T@20.
- The DPP loss requires negative sampling and per-user kernel computation, which the authors approximate.
Experimental setup
Datasets: MovieLens-1M Beauty AMiner Yelp2018
Benchmarks: MovieLens-1M Beauty AMiner Yelp2018
Metrics: Recall@20 NDCG@20 ILAD@20 ILMD@20 T@20
Baselines: BPR-MF LightGCN SGL BUIR SimGCL DGRec
Builds on: SimGCL LightGCN DPP (k-DPP) InfoNCE
Closest papers by meaning
- Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning (AAAI 2025 · similarity 0.89)
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.88)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.87)
- DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation (AAAI 2025 · similarity 0.87)
- Review-Enhanced Hierarchical Contrastive Learning for Recommendation (AAAI 2024 · similarity 0.86)
- Disentangled Contrastive Bundle Recommendation with Conditional Diffusion (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.