Dual-Tower Retrieval

A mainstream recall architecture for handling large-scale candidate sets in industrial recommendation systems. It uses two independent neural networks to encode users and items respectively into dense vectors, measures matching degree through vector similarity, and supports offline pre-computation and approximate nearest neighbor search.