YouTube DNN Candidate Generation

YouTube DNN Candidate Generation (from Google's landmark 2016 paper) formalizes recommendation recall as an extreme multi-class classification problem over millions of video classes; the model embeds variable-length user watch histories via average pooling, concatenating search query tokens, demographic metadata, and the critical 'Example Age' feature (modeling time decay/freshness) through an MLP to generate user vector $u \in \mathbb{R}^d$; trained using Sampled Softmax across video class weights $v_j$, online serving treats the Softmax output weights directly as item embeddings for millisecond ANN dot-product search.

Video streaming recommendations (YouTube, Bilibili), music playlists, and continuous sequence recommendation systems.

Matrix factorization cannot incorporate non-linear temporal dynamics or continuous context signals; YouTube DNN established the paradigm of deep context-rich classification coupled with high-throughput ANN serving.

Explain why the 'Example Age' feature eliminates time bias and enables the model to favor newly uploaded fresh content?

Why does YouTube DNN enforce asymmetric chronological prediction (predicting next watch from past) rather than symmetric CBOW windows?

How does Sampled Softmax approximate full denominator partition functions to enable extreme classification over millions of items?

Why does the learned weight matrix of the Softmax layer mathematically equal item embeddings for online ANN retrieval?

Trace the evolution of watch history aggregation from simple average pooling to target-attention (DIN/SIM) architectures?