Jia Ma, Jun Cheng, Kangyi Lin · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840061
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 (GR) utilizing discrete Semantic IDs (SIDs) has emerged as a promising paradigm for sequential recommendation. However, existing GR paradigms exhibit significant generation flaws: traditional autoregressive (AR) models suffer from strict left-to-right causal constraints leading to error accumulation, while current discrete diffusion adaptations rely on isolated next-item beam search, which causes ''homogeneous stacking'' and lacks explicit list-level coordination. Furthermore, these methods often compromise the interpretable, hierarchical structure of SIDs. To address these limitations, we propose GRID, a novel framework that reformulates generative top-K recommendation as a 2D discrete diffusion inpainting problem. By modeling the recommended item list and their multi-digit hierarchical SIDs along two orthogonal axes, GRID recovers the complete matrix of future items simultaneously via parallel decoding. To ensure structural validity and semantic diversity, we introduce a hierarchical SID construction method alongside a dual-regularized inpainting objective that enforces hierarchy consistency and intra-list uniqueness. Supported by a progressive curriculum training strategy, extensive experiments on three Amazon datasets demonstrate that GRID effectively overcomes homogeneous stacking, fundamentally resolving the limitations of traditional decoding, and achieves state-of-the-art performance across all benchmarks.
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.