{"id":1381,"kind":"story","title":"Two-tower models for search and recommendation: fast retrieval with precomputation, trained by choosing wrong answers","url":"https://zenn.dev/ca_kagglers/articles/two-tower-mechanism","host":"zenn.dev","by":"startupxo","score":0,"comments":0,"time":"2026-11-13T06:31:35.289Z","origin":"curated","language":{"source":"ko","reading":"en","translated":true},"summary":["A two-tower model has one neural network that reads the query or user information and another that reads the items to be searched. It multiplies the number lists (vectors) each produces to score how well they match. Items with higher scores are shown first.","Item-side vectors can be computed and stored in advance, so at search time only the query side needs to be computed. The author explains that this makes it efficient for searching large numbers of candidates.","In training, other correct items in the same batch are reused as wrong answers (negatives). The bias in which popular items are overused as negatives is corrected with their appearance probability (Yi et al. paper), and items that do not appear in the logs are supplemented with random negatives (Yang et al. paper).","On the method of using confusing negatives (hard negatives), the Facebook search paper (Huang et al.) reported that recall was best when using candidates ranked 101 to 500. At deployment, approximate search speeds things up but may miss some candidates."],"discussion":"https://startupxo.com/en/1381/","markdown":"https://startupxo.com/en/1381.md","api":"https://startupxo.com/api/v1/item.json?id=1381&lang=en","text":null,"edited":null,"sources":[{"site":"rss","name":"Zenn","url":"https://zenn.dev/ca_kagglers/articles/two-tower-mechanism","score":null,"comments":null,"observed_at":"2026-11-13T12:27:53Z"}],"thread":[]}