Skip to content

Latest commit

 

History

History
32 lines (18 loc) · 912 Bytes

yt_2016.md

File metadata and controls

32 lines (18 loc) · 912 Bytes

Candidates Generations

Target

  • training as multi-class-classification problem (which video will be watched - class will br > 1M) - will be a problem

Features

video-watch sequence - apply word2vec search - apply word2vec static profile (user)

auto-feature engineering by DNN

Serving

apply ANN

Ranking

Features

  1. same video-watch embedding
    1. why impression and watched?
  2. language embedding