Ad Targeting & pCTR Prediction

The Ad Targeting & pCTR/pCVR Prediction System forms the core revenue backbone of Computational Advertising; it encompasses: 1) The Targeting Engine, which filters active ad campaigns within milliseconds based on advertiser-defined constraints (DMP audience tags, geo-fencing, intent keywords, device profiles); 2) The pCTR/pCVR Neural Estimation Engine, which computes real-time posterior probabilities $p\text{CTR} = P(\text{Click} | u, a, c)$ and $p\text{CVR} = P(\text{Conversion} | \text{Click}, u, a, c)$ using massive sparse deep networks; raw logits undergo rigorous Probability Calibration (Isotonic Regression, Platt Scaling) ensuring predicted probabilities perfectly calibrate with true empirical click frequencies.

Search ads (Google Ads), feed advertising (Meta Ads, TikTok Ads), and display sponsored listings (Amazon Ads).

Computational advertising ties directly to advertiser billing; uncalibrated over-predictions trigger budget burn and advertiser churn, while under-predictions leave ad inventory unsold; rigorous targeting and calibrated ML maintain commercial equilibrium.

Why does recommendation only require relative ordering (high AUC), whereas computational advertising demands absolute calibrated probabilities?

Compare Isotonic Regression vs Platt Scaling in non-parametric calibration and monotonicity guarantees for ad CTR predictions?

How does the Delayed Feedback Model (DFM) mathematically correct for conversions occurring hours/days after the initial click?

Derive the FTRL (Follow The Regularized Leader) streaming algorithm for real-time online ad weight updates with $L_1$ sparsity?

How to enforce strict boundaries between Request-Time features and Post-Impression features to eliminate data leakage in ad logs?