Research map / Image generation / Text-to-image diffusion
Preference Diffusion for 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 PreferDiff, a BPR-derived training objective for diffusion-model sequential recommenders that uses multiple negatives (via their centroid), cosine error, and a generation/preference trade-off, evaluated on Amazon, Steam, ML-1M and Yahoo!R1 benchmarks.
Problem
DM-based recommenders typically train with MSE or standard recommendation losses (BPR, cross entropy), which either fail to model negative items or do not exploit the generative potential of diffusion models for personalized ranking.
Approach
Reformulate BPR as a log-likelihood ranking over diffusion-model likelihoods, derive a variational upper bound that reduces to a denoising loss, swap MSE for cosine error, extend to multiple negatives via their centroid, and mix in the standard generative loss for stability.
Key results
- On Sports and Outdoors, PreferDiff reaches R@5 0.0185 and N@5 0.0147, improving over the best baseline by 19.35% and 16.94% respectively.
- PreferDiff converges in about 35 epochs (8 minutes) versus about 65 epochs (15 minutes) for DreamRec on Amazon Beauty.
- Its text-embedding variant PreferDiff-T gives 2% to 8% gains in-domain over general sequential recommenders in zero-shot transfer experiments.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PreferDiff's preference-aware diffusion objective rather than DreamRec's objective | Diffusion-based sequential recommendation · Sports and Outdoors | Recall@5: 0.0155 -> 0.0185 | Helped |
| Use PreferDiff's preference-aware diffusion objective rather than DreamRec's objective | Diffusion-based sequential recommendation · Beauty | Recall@5: 0.0406 -> 0.0429 | Helped |
| Use PreferDiff's preference-aware diffusion objective rather than DreamRec's objective | Diffusion-based sequential recommendation · Toys and Games | Recall@5: 0.0440 -> 0.0473 | Helped |
| Add the preference-learning term using negative samples (ablation) | PreferDiff · Sports and Outdoors, Beauty, and Toys and Games | Recall@5: 0.0165 -> 0.0185; 0.0415 -> 0.0429; 0.0445 -> 0.0473 | Helped |
| Replace MSE with cosine error (ablation) | PreferDiff · Sports, Beauty, and Toys | Recall@5: 0.0180 -> 0.0185; 0.0393 -> 0.0429; 0.0458 -> 0.0473 | Helped |
| Replace MSE with cosine error and add the preference-learning term (ablation) | PreferDiff · Sports and Outdoors, Beauty, and Toys and Games | Recall@5: 0.0155 -> 0.0185; 0.0406 -> 0.0429; 0.0440 -> 0.0473 | Helped |
| Use text embeddings instead of traditional ID embeddings | PreferDiff · Instruments | Recall@5: 0.1220 vs 0.1213 | No clear effect |
| Use text embeddings instead of traditional ID embeddings | PreferDiff · Tools, CDs, Movies, and Steam | Recall@5: 0.0699 -> 0.0723; 0.0268 -> 0.0295; 0.0306 -> 0.0312; 0.0585 -> 0.0621 | Helped |
| Balance generation and preference learning with λ (ablation) | PreferDiff · Sequential recommendation | Recall@5 | Mixed |
| Train with PreferDiff rather than DreamRec | Amazon Beauty recommendation · Validation training convergence | epochs to convergence: approximately 65 epochs -> approximately 35 epochs | Helped |
Surprising findings
- Performance of DM-based recommenders improves substantially as embedding dimension increases, unlike non-DM recommenders, and PreferDiff degrades sharply at dimension 64.
Limitations
- Performance is highly sensitive to embedding dimension and degrades at the common size of 64.
- Requires careful tuning of the trade-off hyperparameter lambda between generation and preference losses.
Experimental setup
Datasets: Sports and Outdoors Beauty Toys and Games Steam ML-1M Yahoo!R1 Automotive Cell Phones and Accessories Grocery and Gourmet Food Musical Instruments Tools and Home Improvement Office Products
Benchmarks: Sports and Outdoors Beauty Toys and Games Steam Amazon 2018 cross-domain (CDs and Vinyl, Movies and TV)
Metrics: Recall@K NDCG@K
Baselines: GRU4Rec SASRec BERT4Rec CL4SRec TIGER DiffRec DreamRec DiffuRec MoRec LLM2BERT4Rec UniSRec
Builds on: DreamRec BPR DDPM DDIM DPO
Closest papers by meaning
- Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation (NeurIPS 2025 · similarity 0.93)
- On Softmax Direct Preference Optimization for Recommendation (NeurIPS 2024 · similarity 0.90)
- On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders (NeurIPS 2025 · similarity 0.90)
- Learning Multiple User Distributions for Recommendation via Guided Conditional Diffusion (AAAI 2025 · similarity 0.89)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.88)
- Listwise Preference Diffusion Optimization for User Behavior Trajectories Prediction (NeurIPS 2025 · similarity 0.88)
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.