Research map / Image generation / Fast diffusion sampling
Fading to Grow: Growing Preference Ratios via Preference Fading Discrete 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 PreferGrow, a discrete diffusion recommender that fades user preferences by replacing the preferred item via an idempotent fading matrix and grows them back while modeling preference ratios with score entropy loss.
Problem
Diffusion-based recommenders rely on continuous Gaussian noise or simplex-constrained discrete priors, which mismatch the discrete, sparse nature of user preference data and its ranking objective.
Approach
Model relative preference ratios over the discrete item corpus with a discrete diffusion process whose forward step fades the preferred item into alternatives via an idempotent fading matrix (akin to negative sampling), so no Gaussian or Bernoulli prior is needed; reverse process grows preferences from estimated ratios trained with score entropy loss.
Key results
- PreferGrow (Hybrid/Adaptive variants) achieves the best HR@5 and NDCG@5 on Steam (0.0615 HR@5, 0.0406 NDCG@5 for Hybrid), Beauty (0.0420 HR@5) and Sports (0.0207 HR@5) versus the strongest baselines in the table.
- On MovieLens, PreferGrow Hybrid reaches 0.1409 HR@5 and 0.3355 HR@20 at @20, roughly outperforming SASRec and Caser (0.2899 HR@20 for SASRec).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove the general hard negative item (w/o-PointWise) (ablation) | PreferGrow · Steam | recommendation effectiveness | Hurt |
| Replace PairWise preference-ratio modeling with Point-Wise modeling (w/o-PairWise) (ablation) | PreferGrow · Steam | recommendation effectiveness | Hurt |
| Remove non-preference user modeling and personalized enhancement (w/o-Nonpreference) (ablation) | PreferGrow · Steam | recommendation effectiveness | Hurt |
| Increase training duration through the final 50% of training | PreferGrow · recommendation experiments | NDCG@5: about a 5% gain | Helped |
Surprising findings
- Classical sequential recommenders (SASRec, Caser) remain competitive with several diffusion-based baselines on some datasets.
Limitations
- Higher modeling complexity and longer training time than prior diffusion-based recommenders.
- O(N) cost over the item corpus makes it impractical for very large item sets without semantic IDs.
- Rank-r fading matrix and the design space of rank-1 fading matrices are left for future work.
Experimental setup
Datasets: MovieLens Steam Beauty Toys Sports
Benchmarks: MovieLens Steam Beauty Toys Sports
Metrics: HR@5 HR@10 HR@20 NDCG@5 NDCG@10 NDCG@20
Baselines: SASRec Caser GRURec DreamRec PreferDiff DiffRec DDSR
Builds on: SEDD discrete diffusion models (LDR/D3PM) DiffRec DreamRec PreferDiff DDSR SASRec
Closest papers by meaning
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.93)
- 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 0.88)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.87)
- Plug-In Diffusion Model for Sequential Recommendation (AAAI 2024 · similarity 0.87)
- Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback (NeurIPS 2025 · similarity 0.87)
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.