Two-Tower Model

Scoring every item in a catalogue against every request is impossible past a certain size. The two-tower model gets around it by never putting the user and the item into the same network.

Two encoders, one space
One tower encodes the user — history, demographics, context — into a vector. A second encodes the item — identifiers, text, categories — into a vector of the same dimension. The score is their dot product, with no layer anywhere that sees both inputs at once. That restriction is the design, not an oversight: because the item tower depends only on the item, every item embedding can be computed in advance and loaded into an ANN index such as FAISS or ScaNN. At request time only the user tower runs, and retrieval becomes a nearest-neighbour lookup — sub-linear in catalogue size instead of linear.

Training
Training is classification over the catalogue with in-batch negatives or sampled softmax: the other items in the minibatch stand in for the millions not scored. Popular items appear as negatives more often than chance, so the logits are corrected by the sampling probability — skip that and the model systematically suppresses the head of the catalogue.

Where it sits and what it costs
It is a retriever, not a ranker. Without cross-features it cannot model “this user clicks this brand on mobile”, which is what the ranking stage downstream is for. Its other standing cost is freshness: embeddings are only as current as the last index build.