DSSM Two-Tower Retrieval

DSSM Two-Tower Retrieval is an asymmetric neural architecture that decouples user features and item features into two independent sub-networks; the User Tower embeds user demographics, interaction history, and realtime context into vector $u(x) \in \mathbb{R}^d$, while the Item Tower embeds item attributes into $v(y) \in \mathbb{R}^d$, scoring relevance via dot product $\langle u(x), v(y) \rangle$; in production serving, all item embeddings are pre-computed offline and indexed in ANN stores (HNSW/Milvus), requiring only a single User Tower forward pass online followed by millisecond ANN Top-K nearest neighbor search.

Multi-channel recall across e-commerce items, sponsored ad candidates, video recommendations, and cross-modal search.

Classical Matrix Factorization fails on cold-start items and multi-field features, while cross-attention networks cannot be pre-computed offline; Two-Tower networks decouple inference completely, enabling sub-10ms recall over 100M items.

Detail the compute efficiency of In-batch Negative Sampling in Two-Tower training and methods for Popularity Debias correction?

Why does the Two-Tower architecture strictly prohibit early low-level cross-features between user and item inputs?

How to update User Tower embeddings in real-time streaming pipelines to capture instant user interest shifts?

Explain the role and sensitivity of temperature hyperparameter $\tau$ in Two-Tower Softmax cross-entropy loss?

Analyze the optimal sampling mixture of Easy Negatives vs Hard Negatives to prevent embedding collapse and boost discrimination?