Ashish Kumar, Prashant Kumar · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840009
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Generative recommendation represents each item as a semantic ID (SID) sequence and predicts the next item by generating its tokens from user-item interaction history. Recent diffusion-based methods replace autoregressive decoding with masked denoising, enabling bidirectional SID modeling and generation without a fixed left-to-right order. However, masked-diffusion GenRec still depends on two main selection stages: which SID tokens are masked during training and which partial SID candidates are retained during inference. Existing noising strategies estimate token uncertainty from a single probe context, which can make masked views sensitive to context-specific uncertainty, while path-level decoding can retain redundant trajectories that lead to the same partial SID state. We propose S2DGR, a stable-state diffusion framework for generative recommendation. S2DGR introduces Stability-Aware Noising (SAN), which aggregates SID-token uncertainty across multiple probe contexts to construct stable masked training views, and State-Unique Denoising (SUD), which selects inference candidates according to partial SID states rather than only denoising paths. Experiments on multiple benchmark datasets show that S2DGR consistently improves over strong discriminative, generative, and diffusion-based recommendation baselines.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.