Design: Ads CTR Prediction quiz

Question 1

What is the primary purpose of CTR calibration in ad serving systems?

Accepted Answer

To ensure predicted click probabilities accurately match empirical click rates for correct auction pricing

Answer

To maximize the raw predicted CTR score for all ads

Answer

To reduce model training time by simplifying the output layer

Answer

To ensure all ads receive equal click probability predictions

Question 2

How does the Generalized Second Price (GSP) auction work in keyword advertising?

Accepted Answer

The top-ranked ad pays the minimum amount to maintain its rank, based on the next bidder's effective bid

Answer

The top-ranked ad always pays its full bid regardless of other bidders

Answer

All ads pay the same price per click regardless of their bid

Answer

GSP auctions eliminate the need for CTR prediction in ad ranking

Question 3

What is the key difference between CPC and CPM advertising models from an ML perspective?

Accepted Answer

CPC auctions require accurate CTR prediction for ad ranking; CPM auctions optimize impression delivery to audiences

Answer

CPC and CPM have identical ML requirements—only the billing differs

Answer

CPM requires higher CTR prediction accuracy than CPC because impressions are more expensive

Answer

Both CPC and CPM models train on conversion data rather than click data

Question 4

What is the VCG (Vickrey-Clarke-Groves) auction mechanism and when is it used in advertising?

Accepted Answer

VCG is strategy-proof—each bidder's dominant strategy is to bid their true value

Answer

VCG and GSP auctions always produce identical prices and rankings

Answer

VCG always generates higher revenue than GSP auctions for the platform

Answer

VCG auctions do not require CTR prediction because they use raw bids

Question 5

How should sample weighting be applied in CTR model training to correct for data imbalance and selection bias?

Accepted Answer

Apply inverse propensity scoring combined with positive class upweighting to correct class imbalance and selection bias

Answer

Use equal weights for all training samples to avoid introducing any bias

Answer

Always use random undersampling to reduce dataset size before training

Answer

Sample weighting is not necessary when using AUC as the training objective

Question 6

What is normalized cross-entropy (NCE) and why is it preferred over raw log-loss for CTR model evaluation?

Accepted Answer

NCE normalizes log-loss by a background CTR baseline, enabling fair comparison across datasets with different base rates

Answer

NCE measures the total revenue generated per ad impression

Answer

NCE is the same as AUC-ROC and measures ranking quality only

Answer

NCE is only applicable to CPA campaigns, not CPC campaigns

Test yourself