Research map / Image generation / Fast diffusion sampling
Disentangled Contrastive Bundle Recommendation with Conditional Diffusion
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 DCBR, a bundle recommender that denoises the user-bundle graph with a conditional diffusion model (with a bundle latent consistency constraint), fuses triple-view graph embeddings, and applies dual-level (inter/intra-view) disentangled contrastive learning.
Problem
Bundle recommendation suffers from noisy user-bundle interactions and sparsity, and existing SSL methods that contrast final fused representations introduce semantic noise across views.
Approach
A conditional bundle diffusion model denoises the user-bundle graph, regularized by a bundle latent consistency loss so denoising does not erase genuine interactions; a dual-level contrastive loss contrasts pre-fusion view representations both across views (inter-view) and within each view (intra-view).
Key results
- On MealRec+ H, DCBR reaches Recall@20 0.5261 and NDCG@20 0.3483, versus 0.4211 and 0.2693 for the second-best MultiCBR (reported improvements of 24.93% and 29.34%).
- On iFashion, DCBR achieves Recall@20 0.1633 and NDCG@20 0.1370, improving NDCG@20 by 13.88% over the best baseline.
- Ablation removing the DDCL loss drops performance sharply (e.g., MealRec+ H Recall@10 from 0.4113 to 0.0174).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BLCC loss to conditional bundle diffusion (ablation) | DCBR · MealRec+H, MealRec+L, iFashion | R@20: 0.5025 -> 0.5261; 0.3597 -> 0.3611; 0.1612 -> 0.1633 | Helped |
| Add conditional bundle diffusion denoising (CBDM) (ablation) | DCBR · MealRec+H, MealRec+L, iFashion | R@20: 0.4731 -> 0.5261; 0.3464 -> 0.3611; 0.1495 -> 0.1633 | Helped |
| Add inter-view contrastive learning (ablation) | DCBR · MealRec+H, MealRec+L, iFashion | R@20: 0.2856 -> 0.5261; 0.2616 -> 0.3611; 0.1462 -> 0.1633 | Helped |
| Add intra-view contrastive learning (ablation) | DCBR · MealRec+H, MealRec+L, iFashion | R@20: 0.5070 -> 0.5261; 0.3375 -> 0.3611; 0.1611 -> 0.1633 | Helped |
| Add dual-level disentangled contrastive learning (ablation) | DCBR · MealRec+H, MealRec+L, iFashion | R@20: 0.0410 -> 0.5261; 0.0814 -> 0.3611; 0.0404 -> 0.1633 | Helped |
Surprising findings
- Removing the entire DDCL contrastive loss causes a drastic collapse (e.g., Recall@10 on MealRec+ H falls from 0.4113 to 0.0174), far more than removing either the inter-view or intra-view term alone.
Limitations
- Contrastive loss time complexity grows quadratically with batch size B (O(B^2 d)).
- Hyperparameters (noise intensities, fusion weights, temperature, loss weights) are tuned by grid search, with validation-based selection.
Experimental setup
Datasets: MealRec+ H MealRec+ L iFashion
Benchmarks: MealRec+ H MealRec+ L iFashion
Metrics: Recall@10 Recall@20 NDCG@10 NDCG@20
Baselines: Pop MF-BPR NGCF LightGCN SGL SimGCL XSimGCL BIGCF BGCN UHBR CrossCBR DSCBR
Builds on: DiffRec DDPM SimGCL MultiCBR BPR
Closest papers by meaning
- DisCo: Graph-Based Disentangled Contrastive Learning for Cold-Start Cross-Domain Recommendation (AAAI 2025 · similarity 0.88)
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.88)
- Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning (AAAI 2025 · similarity 0.87)
- Learning Multiple User Distributions for Recommendation via Guided Conditional Diffusion (AAAI 2025 · similarity 0.86)
- DivGCL: A Graph Contrastive Learning Model for Diverse Recommendation (AAAI 2025 · similarity 0.85)
- Disentangled Modeling of Preferences and Social Influence for Group Recommendation (AAAI 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.