Reproduce “Quantizing Intent: Cross-Domain Semantic IDs from Organic Activity for Industrial Ranking” 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
Julie Choi, Haoran Ye, Zhiwei Ding, Bo Long, Benjamin Zelditch, Arpita Vats
The paper from LinkedIn develops methods to create compact, discrete user representations (Semantic IDs) from organic platform activity, integrating them into an industrial ads click-through rate prediction model. This approach improves ads ranking performance, especially for cold-start users, by leveraging rich cross-domain behavioral signals while significantly reducing embedding storage costs.
Key insights
- The "Behavioral Activity Richness Principle" establishes a monotonic relationship between the amount of behavioral activity encoded in a source embedding and its downstream transfer quality for ads ranking.
- RQ-FSQ effectively preserves both global embedding geometry and per-dimension fine structure, achieving substantial storage reductions (30x to 280x) with equivalent or improved AUC compared to dense float embeddings.
- Cross-domain viewer SIDs, combined with end-to-end fine-tuning under the target task objective, serve as an effective "behavioral cold-start bridge" for users with sparse ad interaction history.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.