Research map / Image generation / Fast diffusion sampling
Enhancing Sequential Recommendation with Global 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 GlobalDiff, a plug-and-play diffusion model stacked on frozen sequential recommenders to restore global item-item structure and ensemble it with the backbone's sequential predictions.
Problem
Sequential recommendation models are misled by inconsistent, externally driven items in the local interaction sequence, and ignore global non-sequential item structure.
Approach
Build masked-item training samples from an item-item co-occurrence similarity matrix with positional re-weighting and Bernoulli sampling, train a diffusion model whose denoiser is guided by a frozen backbone's context representation, and combine the diffusion output with the backbone's softmax via Borda-count voting at inference.
Key results
- GlobalDiff improves backbone sequential models by an average of 9.67% across datasets and metrics (next-item prediction).
- On SASRec, GlobalDiff gives an average gain of 17.06%; on CBiT, a 5.10% average gain.
- Long-term prediction: average improvement of 9.40% in HR@20 across three backbones and all datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GlobalDiff to CBiT | CBiT · ML-1M | HR@10: 0.1542 -> 0.1638 | Helped |
| Add GlobalDiff to CBiT | CBiT · KuaiRec | HR@10: 0.1590 -> 0.1672 | Helped |
| Add GlobalDiff to CBiT | CBiT · Beauty | HR@10: 0.0476 -> 0.0510 | Helped |
| Add mean-pooled sequence representation to the target representation in the reverse approximator (ablation) | GlobalDiff reverse approximator · ML-1M | performance: +30% | Helped |
| Remove the guided representation from the reverse approximator (ablation) | GlobalDiff reverse approximator · ML-1M | performance: 81% | Hurt |
| Use only the target-item representation, omitting the mean-pooled sequence representation (ablation) | GlobalDiff reverse approximator · ML-1M | performance: +30% | Hurt |
| Combine diffusion and sequential predictions with voting (ablation) | GlobalDiff inference · ML-1M | HR@20: 0.237 -> 0.247 | Helped |
| Increase diffusion steps from 20 to 100 or 200 (ablation) | GlobalDiff · ML-1M | HR@20: less than 0.005 | No clear effect |
| Use cluster-item similarity with K=100 instead of pair-wise similarity (ablation) | CBiT+GlobalDiff · Beauty | HR@20: 0.0585 -> 0.0580 | No clear effect |
| Use cluster-item similarity with K=500 instead of pair-wise similarity (ablation) | CBiT+GlobalDiff · Beauty | HR@20: 0.0585 -> 0.0618 | Helped |
| Use cluster-item similarity with K=1000 instead of pair-wise similarity (ablation) | CBiT+GlobalDiff · Beauty | HR@20: 0.0585 -> 0.0625 | Helped |
Surprising findings
- DiffuRec, a diffusion model used as the main predictor, performs worse than backbones BERT4Rec and CBiT.
- Removing the guided sequential representation from the reverse approximator causes the largest performance drop, with the poorest results among variants.
Limitations
- Pairwise item-item similarity aggregation is O(M^2), which is costly for large item universes without the clustering alternative.
- Gains on some metrics are marginal or negative (e.g., SRGNN HR@20 on ML-1M is -0.69%).
Experimental setup
Datasets: MovieLens 1M Amazon-Beauty KuaiRec
Benchmarks: MovieLens 1M Amazon-Beauty KuaiRec
Metrics: HR@10 HR@20 NDCG@10 NDCG@20 Recall@K
Baselines: Gru4Rec Caser LightSANS DuoRec CL4Rec AC TSR DiffuRec
Builds on: DDPM CBiT BERT4Rec SASRec Borda count
Closest papers by meaning
- Enhancing Diffusion Model with Auxiliary Information Mining-Exploration and Efficient Sampling Mechanism for Sequential Recommendation (AAAI 2025 · similarity 0.89)
- Breaking Determinism: Fuzzy Modeling of Sequential Recommendation Using Discrete State Space Diffusion Model (NeurIPS 2024 · similarity 0.87)
- SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.87)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.87)
- LLMEmb: Large Language Model Can Be a Good Embedding Generator for Sequential Recommendation (AAAI 2025 · similarity 0.86)
- SIGMA: Selective Gated Mamba for Sequential Recommendation (AAAI 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.