Reproduce “Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junting Wang, Xinrui He, Yunzhe Li, Hari Sundaram
The paper introduces Item-Supported Decoding (ISD), which uses an external user-history item ranking to help semantic-ID prefixes survive beam pruning and to reorder generated items. In controlled tests, semantic IDs changed 38.4% on average after field reordering, while top-10 neighbor overlap was 32.2%. On Amazon Baby, ISD raised NDCG@10 from 0.0295 to 0.0387, but gains varied by metric and setting.
Key insights
- Exact Semantic ID (SID) assignments are unstable, with 14.0% to 73.5% of items receiving different SIDs for semantically equivalent descriptions, and SIDs often fail to preserve fine-grained local item structure.
- Autoregressive SID generation, through beam search, prematurely removes a substantial fraction (e.g., 24.4% to 35.4% for TIGER) of highly-ranked target items.
- Incorporating item-level evidence from an independent ranker into beam search can drastically increase the availability of relevant targets, improving retention from 35.4% to 88.3% for TIGER (Office) after the first token.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.