Research map / Image generation / Fast diffusion sampling
Learning Multiple User Distributions for Recommendation via Guided 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 GCDR, a guided conditional diffusion recommender that learns multiple user distributions (conditioned on long-term user embedding, guided by short-term sequence and category preference) and trains with user, recommendation, and diffusion losses.
Problem
Embedding-based recommenders use a single fixed user embedding and struggle with user uncertainty and diverse interests; existing user-distribution methods use only a single distribution per user, blurring semantic (e.g., category) boundaries.
Approach
Model each user as multiple distributions over embeddings of interacted items: inter-user distributions come from conditional diffusion on a user-specific embedding, and intra-user distributions come from classifier-free guided diffusion using short-term interest (Transformer over recent items) and category preference, with category-specific sampling at inference.
Key results
- GCDR achieves best Recall@20 on all four datasets, e.g., 0.1319 on MovieLens vs. 0.1265 for DiffuRec (runner-up).
- On Yelp, GCDR reaches Recall@20 0.1552 and NDCG@20 0.0809 vs. 0.1371 / 0.0711 for DiffuRec.
- Ablations show removing the recommendation loss collapses performance (MovieLens Recall 0.0049 vs. 0.1319).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use guided conditional diffusion to learn multiple user distributions | GCDR · MovieLens | Recall@20: 0.1265 -> 0.1319 | Helped |
| use guided conditional diffusion to learn multiple user distributions | GCDR · KuaiRec | Recall@20: 0.0617 -> 0.0733 | Helped |
| use guided conditional diffusion to learn multiple user distributions | GCDR · Yelp | Recall@20: 0.1371 -> 0.1552 | Helped |
| use guided conditional diffusion to learn multiple user distributions | GCDR · Books | Recall@20: 0.1273 -> 0.1363 | Helped |
| add long-term interest cl (ablation) | GCDR · MovieLens | Recall@20: 0.1209 -> 0.1319 | Helped |
| add short-term interest gs (ablation) | GCDR · MovieLens | Recall@20: 0.1019 -> 0.1319 | Helped |
| add category preference gc (ablation) | GCDR · MovieLens | Recall@20: 0.1252 -> 0.1319 | Helped |
| add user loss Luser (ablation) | GCDR · MovieLens | Recall@20: 0.1242 -> 0.1319 | Helped |
| add recommendation loss Lrec (ablation) | GCDR · MovieLens | Recall@20: 0.0049 -> 0.1319 | Helped |
| add diffusion loss Ldiff (ablation) | GCDR · MovieLens | Recall@20: 0.0581 -> 0.1319 | Helped |
| add category-specific distributions during inference (ablation) | GCDR · MovieLens | Recall@20: 0.1265 -> 0.1319 | Helped |
Surprising findings
- Removing the recommendation loss causes a dramatic collapse in performance (MovieLens Recall drops from 0.1319 to 0.0049).
Limitations
- Category of the target item is not used during training, left for future work.
- Performance sensitive to hyperparameters such as the recommendation-loss threshold tau, the number of categories L, and the xt coefficient delta.
Experimental setup
Datasets: MovieLens KuaiRec Yelp Books
Benchmarks: MovieLens KuaiRec Yelp Books
Metrics: Recall@20 HR@20 NDCG@20
Baselines: MF NCF NGCF LightGCN GRU4REC SASRec STOSA HCUR CBox4CR LCD-UC DiffRec DiffuRec
Builds on: DDPM DiffRec DreamRec classifier-free guidance DiffuRec
Closest papers by meaning
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.89)
- Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation (NeurIPS 2025 · similarity 0.88)
- Enhancing Diffusion Model with Auxiliary Information Mining-Exploration and Efficient Sampling Mechanism for Sequential Recommendation (AAAI 2025 · similarity 0.87)
- On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders (NeurIPS 2025 · similarity 0.86)
- Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling (AAAI 2025 · similarity 0.86)
- Density-based User Representation using Gaussian Process Regression for Multi-interest Personalized Retrieval (NeurIPS 2024 · similarity 0.86)
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.