Рекомендательные модели с двумя башнями

Recommendation systems often separate candidate retrieval from ranking. Retrieval narrows a large catalog to a manageable set of potentially relevant items. A two-tower model uses one encoder for the user or query and another for candidate items. Each tower transforms its own available features into a vector in a shared embedding space. A similarity function, often a dot product or cosine score, ranks candidate pairs. The two encoders can be run independently. Item embeddings may be calculated offline and stored in a nearest-neighbor index. At request time, the user tower creates a query vector and the index returns a shortlist. This separation supports large catalogs more efficiently than evaluating a deep joint model against every item. TensorFlow Recommenders describes this architecture as retrieval with query and candidate models. The independent structure creates a tradeoff. It scales retrieval, but it cannot directly model arbitrary pairwise interactions that require seeing the user and item together, unless those interactions are represented in the embedding training objective or features. A downstream ranker can combine retrieved candidates with richer context, cross-features and operational constraints. Retrieval recall matters: a ranker cannot select an item that was never retrieved. Training often uses positive interactions and sampled or in-batch negatives to shape the embedding space. The negative sampling strategy and exposure logs affect what the model learns. Evaluation should measure recall or ranking among a realistic candidate pool and check cold-start performance. Monitor embedding freshness, index refreshes and feature-version compatibility. A high similarity is a retrieval score, not a calibrated probability or proof of user preference. Eligibility, safety, diversity and policy constraints should be applied in the system design. The architecture makes large-scale search practical, while final user experience depends on candidate coverage, ranking quality and evaluation beyond offline similarity.