You already know: logistic regression and log loss (D3), embeddings for sparse identifiers (D4), calibration (D3) and, from the recommendation case, staged retrieval and ranking. This case adds prediction that is priced: the predicted click probability is multiplied by an advertiser's bid to run an auction, so a calibration error is a pricing error. The design questions: why calibration matters more than ranking, how to stay fresh when ads change hourly, how to train on billions of rare events, and how to experiment when advertisers' budgets couple the test arms.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud; prices are illustrative.
1. The prompt, and the questions that change the design
A mid-level answer: "train a deep model on impressions and clicks and rank ads by predicted CTR". A staff candidate first asks how the prediction is used in the auction, the scale and latency budget, which labels exist and how delayed they are, and what guardrails apply for users and advertisers. Those answers produce calibrated probabilities, a light-then-heavy funnel, continuous training with corrected downsampling, and ad load and policy limits.
Assumptions you state: