YouTube DNN: A deep user tower for candidate generation¶

Trains history pooling, nonlinear user tower and item embeddings, not private logs or production sampled softmax.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	Google / YouTube.
First public release	2016-09-15
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`youtube-dnn`
Local reproduction code	`src/auto_research/reproductions/youtube_dnn/`

Paper method and evidence¶

Separate fast candidate generation from fine ranking. Aggregate watch-history embeddings and other heterogeneous context with a deep nonlinear user tower. Predict the next video as large-scale multiclass classification and retrieve by nearest neighbor at serving.

flowchart LR
  A[Watch history and context] --> B[Deep user tower]
  B[Deep user tower] --> C[Nearest-neighbor video retrieval]

Original paper figure¶

The original figure shows candidate generation and ranking. From the original paper; copyright remains with the authors.

Core equation¶

\[ P(v=i\mid U)=\frac{\exp(v_i^\top u)}{\sum_j\exp(v_j^\top u)}. \]

Reported results¶

The paper describes online architecture and offline comparisons but no quantified online lift meeting this repository's inclusion threshold. User-approved classic exception.

Local reproduction¶

Complete MovieLens-100K, same item embeddings, negative samples, budget and full catalog. History mean-pooling two-tower vs nonlinear tower: three-seed NDCG@10 0.02854→0.02666 (-6.61%). Retain the negative result. Metrics. Run auto-research reproduce --paper youtube-dnn --dataset-dir data.