Research map / Probabilistic and scientific modeling / Molecules and proteins
Breaking Determinism: Fuzzy Modeling of Sequential Recommendation Using Discrete State Space Diffusion Model
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 DDSR, a discrete state-space diffusion model for sequential recommendation that builds fuzzy sets of interaction sequences via structured transitions over semantic IDs (PQ/RQ-VAE) and trains a Transformer to predict the next item.
Problem
Sequential recommendation assumes deterministic user behavior and conventional models fail to capture randomness; continuous-space diffusion recommenders (DiffuRec, DreamRec) relax discrete interactions into noisy continuous embeddings, causing information loss.
Approach
Apply discrete-state diffusion directly to the item-sequence (blurring the input 'cause' rather than the target), with uniform or Gaussian-distance importance transition matrices over compact semantic-ID codebooks, and a Transformer approximator trained with cross-entropy, justified by an information-diffusion completeness argument.
Key results
- DDSR_T reaches R@10 0.1207 and N@10 0.0663 on Scientific, versus 0.1145 and 0.0594 for the best baseline DiffuRec (ID).
- On Office, DDSR R@10 is 0.1138 vs 0.1090 for VQ-Rec (+4.40%); on Online Retail R@10 is 0.1687 vs 0.1532 for VQ-Rec (+10.12%).
- Ablation: removing diffusion (w/o diffusion) drops Scientific R@10 from 0.1207 to 0.1126; Random ID drops R@10 to 0.0554, below SASRec.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DDSR with text semantic IDs and discrete diffusion | Sequential recommendation · Scientific | Recall@10: 0.1145 -> 0.1207 | Helped |
| use DDSR with text semantic IDs and discrete diffusion | Sequential recommendation · Office | NDCG@10: 0.0714 -> 0.0768 | Helped |
| use DDSR with text semantic IDs and discrete diffusion | Sequential recommendation · Online Retail | NDCG@10: 0.0713 -> 0.0876 | Helped |
| add discrete diffusion to PQ semantic IDs (ablation) | DDSR · Scientific | Recall@10: 0.1126 -> 0.1207 | Helped |
| add discrete diffusion to PQ semantic IDs (ablation) | DDSR · Office | Recall@10: 0.1076 -> 0.1138 | Helped |
| add discrete diffusion to PQ semantic IDs (ablation) | DDSR · Online Retail | Recall@10: 0.1549 -> 0.1517 | Hurt |
| use importance rather than uniform transition (ablation) | DDSR · Online Retail | Recall@10: 0.1517 -> 0.1687 | Helped |
| use importance rather than uniform transition (ablation) | DDSR · Scientific | Recall@10: 0.1207 -> 0.1192 | Hurt |
| replace PQ IDs with RQ-VAE IDs (ablation) | DDSR · Online Retail | Recall@10: 0.1207 -> 0.1081 | Hurt |
| replace PQ semantic IDs with random IDs (ablation) | DDSR · Scientific | Recall@10: 0.1207 -> 0.0554 | Hurt |
| replace PQ semantic IDs with random IDs (ablation) | DDSR · Office | Recall@10: 0.1138 -> 0.0548 | Hurt |
| replace PQ semantic IDs with random IDs (ablation) | DDSR · Online Retail | Recall@10: 0.1687 -> 0.0428 | Hurt |
Surprising findings
- Random semantic IDs underperform the SASRec baseline, indicating that semantic content rather than code structure drives gains.
- DDSR is slower than UniSRec in training and evaluation and uses more GPU memory than UniSRec, despite the claimed efficiency of inference in k-step jumps.
Limitations
- Higher training and inference cost than non-diffusion baselines and DiffuRec's efficiency gains are limited.
- TIGER reimplemented without released code, so its results may deviate from the original.
- Theoretical completeness argument is loosely stated and the proof is hard to follow.
Experimental setup
Datasets: Amazon Scientific Amazon Office Online Retail
Benchmarks: Amazon Scientific Amazon Office Online Retail
Metrics: Recall@10 NDCG@10 Recall@50 NDCG@50
Baselines: SASRec BERT4Rec UniSRec VQ-Rec TIGER ACVAE DiffuRec DreamRec
Builds on: D3PM VQ-Rec TIGER DiffuRec DDPM
Closest papers by meaning
- Enhancing Diffusion Model with Auxiliary Information Mining-Exploration and Efficient Sampling Mechanism for Sequential Recommendation (AAAI 2025 · similarity 0.87)
- Enhancing Sequential Recommendation with Global Diffusion (AAAI 2025 · similarity 0.87)
- Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling (AAAI 2025 · similarity 0.85)
- Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation (NeurIPS 2025 · similarity 0.85)
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.85)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.84)
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.