Two-tower recsys — architecture and use.

hard

Answer

$User tower: NN encoding user features / history \to user embedding.$
$Item tower: NN encoding item features \to item embedding.$
$Score = cosine or dot product.$
$Trained with softmax over batch (in-batch negatives).$
$Deployment: precompute item embeddings, index with ANN; user embedding computed at query time \to real-time retrieval.$
$Standard in YouTube, TikTok, Instagram, LinkedIn retrieval-stage recsys.$

Check yourself — multiple choice

Same as MF
User NN + item NN → dot-product score; in-batch softmax negatives; precomputed item ANN + on-the-fly user embed → real-time retrieval; TikTok/YouTube standard
Random
Not real

Two-tower: user + item encoders + ANN on items; standard retrieval-stage recsys.

#applications

214 interview questions in this topic.