A retrieval model with two neural networks, one that embeds the user or query and one that embeds the item, trained so that the dot product of the two embeddings scores relevance. Item embeddings are computed in advance, so millions of candidates can be searched quickly with nearest-neighbour search.
Card 1 of 228. Answer: A retrieval model with two neural networks, one that embeds the user or query and one that embeds the item, trained so that the dot product of the two embeddings scores relevance. Item embeddings are computed in advance, so millions of candidates can be searched quickly with nearest-neighbour search.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
The user tower takes user features and recent history; the item tower takes item features such as ID, category and text. Both output vectors of the same size, and training on logged interactions pushes the dot product up for pairs that interacted and down for negatives, often the other items in the same batch.
Example: a video platform with 10 million videos runs the item tower offline and stores the vectors in an approximate nearest-neighbour index (FAISS or ScaNN, for example). At request time it runs the user tower once, retrieves the 500 videos with the highest dot product in a few milliseconds, and hands them to a heavier ranking model that scores each user-video pair with richer features. This retrieve-then-rank design is the standard answer in recommender system design interviews.
The trade-off: the two towers only meet at the dot product, so the model can't learn interactions between user and item features. That is why it retrieves candidates rather than producing the final ranking. In-batch negatives over-penalize popular items, which Yi et al. corrected for at YouTube (RecSys 2019). A plus: the item tower can embed a new item from its features, which eases cold start.
