Reproduce “Exploring Bottom-Up Clustering for Creating Semantic IDs” 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
Leah Woldemariam, Sudhanshu Garg, Taha Belkhouja, Charles Kim-Yip, Ali Sahami
The authors propose bottom-up product coding for generative recommendation, where a model generates an item identifier from a user’s history. The method clusters similar product embeddings finely, gives each item a unique final code, then merges clusters into broader groups. Prediction improves on a 5.8-million-product catalog and Beauty, but not Sports & Outdoors; the comparison does not establish causation.
Key insights
- Embedding discretization order matters: building local neighborhoods before coarsening yields better-separated fine-grained groups than constructing the hierarchy primarily from coarse residual representations.
- Uniqueness can be enforced during identifier construction rather than repaired afterward with arbitrary appended codes, preserving a clearer connection between prefixes and semantic neighborhoods.
- Improved clustering quality does not guarantee universal downstream gains: results depend on the catalog, and category coherence and cosine similarity favor different organizational properties.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.