Feed Ads CTR System: Training Data, Bias, Calibration and Embeddings
Problem statement
How would you improve ad ranking in an image-based social feed without hurting the organic experience or long-term engagement? Ad clicks are sparse and many creatives are new, while related organic interactions are much denser. The auction ranks bids multiplied by predicted click probability, so those probabilities must be calibrated.
Fix the training data, correct auction probabilities, and choose embeddings for sparse ads.
Show hint
Ask what probability the auction actually needs, and which logged events let you estimate it for an ad that was shown.
Relevant roles: ML Engineer, Applied Scientist and Data Scientist.