YouTube RecSys Case Study (the two papers everyone cites)

Knowing these two papers cold lets you anchor almost any recsys interview, especially at Google.

Paper 1 — “Deep Neural Networks for YouTube Recommendations” (Covington et al., 2016)

Defined the modern two-stage shape (The Two-Stage RecSys Funnel).

Candidate generation as extreme multiclass classification. “Which of millions of videos will this user watch next?” — a softmax over the corpus, trained with sampled softmax. User vector = MLP over [average of watched-video embeddings, search-token embeddings, demographics, example age]. Serving = user vector + ANN over video vectors — i.e., a two-tower in modern terms.

Details interviewers fish for:

Example age. Models trained on weeks of logs learn the average popularity of a video over the window, but users prefer fresh content; the model can’t tell “popular now” from “was popular.” Fix: feed the age of the training example as a feature (set to 0 at serving) → the model learns the time-dependence explicitly and serving at “age 0” reflects current preference. A classic distribution-shift fix.
Predict the next watch, not a held-out random watch — random held-out leaks future info into features (you’d train on context that includes events after the label). Always make train/test splits temporal, per example.
Per-user example caps so whales don’t dominate the loss.
Withhold-the-last-search-token style tricks to prevent the model from just parroting the immediately preceding query.

Ranking with weighted logistic regression for watch time. Positives weighted by watch time; with low click probability, the learned odds $e^{s}$ ≈ expected watch time → rank by $e^s$. (Explained in Loss Functions Field Guide §9.) The general lesson: you can smuggle a regression target into a classification model via example weights.

Paper 2 — “Recommending What Video to Watch Next” (Zhao et al., 2019)

The modern ranking stack:

Multi-task MMoE: engagement tasks (click, watch time) + satisfaction tasks (like, dismissal, survey responses), mixed by per-task gates; final score = weighted combination (weights = policy).
Position/selection bias correction with a “shallow tower”: a small side-model takes position (and device) and produces a bias logit added to the main logit during training only; at serving, position is dropped (or set to a fixed value). The main tower is thus trained to explain engagement net of position. This is the production-grade answer to position bias — cheap, effective, widely copied.

The classic follow-up questions and answers

“Why not rank by pCTR alone?” — clickbait optimization; watch time / satisfaction heads + negative-feedback heads (dismissals) counterbalance.
“How does the system avoid filter bubbles / feedback loops?” — exploration traffic, diversity re-ranking, satisfaction objectives; log-and-learn loops discussed in Monitoring, Drift, and Feedback Loops.
“How would you add Shorts/long-video tradeoffs?” — per-format value models; calibrate dwell across formats (a 30s Short ≠ 30s of a documentary) — normalize by format-conditional expectations.