Research map / Image generation / Fast diffusion sampling
Plug-In Diffusion Model for Sequential 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 PDRec, a model-agnostic plug-in framework that uses a time-interval diffusion recommender to reweight historical behaviors, augment positives from top-ranked unobserved items, and sample safe negatives for sequential recommenders.
Problem
Existing diffusion-based recommenders use only the highest-scored item from the diffusion output and ignore time intervals and generalized preferences over all items, leaving data sparsity unaddressed and failing to integrate with sequential recommenders.
Approach
Treat a pre-trained time-interval diffusion model (TI-DiffRec) as a plugin that supplies diffusion-based preferences over all items, which are used for three strategies: Historical Behavior Reweighting (HBR), Diffusion-based Positive Augmentation (DPA) via self-distillation, and Noise-free Negative Sampling (NNS).
Key results
- PDRec with SASRec improves N@1 on the Toy dataset from 0.1095 to 0.1247 (13.88% relative gain).
- PDRec with CL4SRec improves N@1 on Toy from 0.1125 to 0.1254 (11.47%) and HR@10 on Toy from 0.3195 to 0.3595.
- On cross-domain mixed settings PDRec improves over T-DiffRec/TI-DiffRec by up to 38.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PDRec to SASRec | SASRec · Toy, Game, Book, and Music | N@10: Toy: 0.2020 -> 0.2286; Game: 0.3682 -> 0.3753; Book: 0.5272 -> 0.5393; Music: 0.5503 -> 0.5620 | Helped |
| Add HBR (ablation) | SASRec · Toy and Game | N@10: Toy: 0.2261 -> 0.2311; Game: 0.3905 -> 0.3963 | Helped |
| Add NNS (ablation) | SASRec + HBR · Toy and Game | N@10: Toy: 0.2311 -> 0.2323; Game: 0.3963 -> 0.3975 | Helped |
| Add DPA (ablation) | SASRec + HBR + NNS · Toy and Game | N@10: Toy: 0.2323 -> 0.2348; Game: 0.3975 -> 0.3992 | Helped |
| Use TI-DiffRec instead of replacing it with SASRec (ablation) | PDRec · Sequential recommendation | recommendation performance | Helped |
| Add PDRec to SASRec | SASRec · Toy → Game and Game → Toy | N@1: Toy → Game: 0.1267 -> 0.1302; Game → Toy: 0.2273 -> 0.2363 | Helped |
Surprising findings
- Pure DM-based recommenders (T-DiffRec, TI-DiffRec) remain inferior to standard SOTA sequential recommenders such as SASRec.
- On the denser Douban datasets the gains from PDRec over backbones are smaller (around 1-2% for SASRec and CL4SRec).
Limitations
- Gains on denser datasets are smaller than on sparse Amazon datasets.
- Many hyperparameters (omega_m, m, n, omega_d, HBR weights) are dataset-specific and tuned per dataset.
- Evaluation uses 99 sampled negatives per positive, which may not reflect full-corpus ranking.
Experimental setup
Datasets: Amazon Toys and Games Amazon Video Games Douban Books Douban Music
Benchmarks: Toy Game Book Music
Metrics: NDCG@k HR@k AUC
Baselines: T-DiffRec TI-DiffRec GRU4Rec SASRec CL4SRec
Builds on: DiffRec T-DiffRec SASRec CL4SRec
Closest papers by meaning
- Enhancing Diffusion Model with Auxiliary Information Mining-Exploration and Efficient Sampling Mechanism for Sequential Recommendation (AAAI 2025 · similarity 0.89)
- Curriculum Conditioned Diffusion for Multimodal Recommendation (AAAI 2025 · similarity 0.88)
- Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation (NeurIPS 2025 · similarity 0.87)
- On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders (NeurIPS 2025 · similarity 0.87)
- Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling (AAAI 2025 · similarity 0.86)
- Preference Diffusion for Recommendation (ICLR 2025 · 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.