Research map / Image generation / Fast diffusion sampling
Enhancing Diffusion Model with Auxiliary Information Mining-Exploration and Efficient Sampling Mechanism 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 DAE4Rec, a conditional diffusion recommender that enhances item embeddings with a graph auto-encoder (norm-based alignment loss) and accelerates sampling via a non-Markov 'skip-step' reverse process with a learned compensator.
Problem
Sequential recommendation with diffusion models suffers from poorly structured item/condition representations (high coupling) and slow inference during sampling.
Approach
Injects global item-transition structure into diffusion embeddings via a GAE-reconstructed item-item graph aligned through a matrix-norm loss, and uses a non-Markov DDIM-style skip-step sampler with a compensator network to correct the resulting bias in the final denoising step.
Key results
- DAE4Rec achieves HR@20 of 48.98% on Foursquare vs. 27.25% for DreamRec and 48.65% for SASRec, the best across all three datasets.
- On Gowalla, DAE4Rec reaches NDCG@20 of 25.50% vs. 25.18% for DreamRec (HR@20 28.32% vs. 25.18%-related baselines).
- With sampling steps S set to half of T, the authors report under 4% overall performance degradation while substantially reducing inference time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AIME auxiliary information mining and exploration (ablation) | DAE4Rec · Foursquare | HR@10 (%): 27.53 -> 39.88 | Helped |
| Use the reconstructed graph target instead of the ground-truth graph target (ablation) | DAE4Rec · Foursquare | HR@10 (%): 31.71 -> 39.88 | Helped |
| Add the sampling compensator (ablation) | DAE4Rec · Foursquare | HR@10 (%): 37.10 -> 39.88 | Helped |
| Use DAE4Rec instead of SASRec | sequential recommendation · Foursquare | HR@10 (%): 38.93 -> 39.88 | Helped |
| Use DAE4Rec instead of DreamRec | sequential recommendation · Gowalla | HR@10 (%): 21.59 -> 23.99 | Helped |
| Use DAE4Rec instead of DreamRec | sequential recommendation · Retailrocket | HR@10 (%): 38.19 -> 39.46 | Helped |
Surprising findings
- Plain SASRec is competitive with or better than several generative baselines on Foursquare, and the DM-based DreamRec underperforms it there on HR@20.
Limitations
- Sampling-step reduction causes some performance degradation, larger on longer sequences.
- Performance sensitive to the AIME loss weight delta; too large values hurt generation quality.
Experimental setup
Datasets: Foursquare Gowalla Retailrocket
Benchmarks: Foursquare Gowalla Retailrocket
Metrics: HR@10 HR@20 NDCG@10 NDCG@20
Baselines: Caser SASRec STOSA ACVAE ContrastVAE DiffRec DreamRec
Builds on: DreamRec DDIM DDPM DiffRec
Closest papers by meaning
- Enhancing Sequential Recommendation with Global Diffusion (AAAI 2025 · similarity 0.89)
- Plug-In Diffusion Model for Sequential Recommendation (AAAI 2024 · similarity 0.89)
- On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders (NeurIPS 2025 · similarity 0.88)
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.88)
- Breaking Determinism: Fuzzy Modeling of Sequential Recommendation Using Discrete State Space Diffusion Model (NeurIPS 2024 · similarity 0.87)
- Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling (AAAI 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.