Ad click prediction

How this question breaks down

1. 1CTR modelCore17 minPredict the probability that a sponsored card is clicked, from logged impressions, sparse IDs and context, accurately enough that an auction can multiply it by a bid. Measured by normalised entropy and calibration, not AUC alone.
2. 2Sparse featuresCore23 minAn ad model's size is mostly embedding rows for IDs and ID pairs, and most of those IDs are seen once. Decide which IDs get a row, when rows expire, how precise they are, and how the tables are split across machines for training and for a 20 ms serving slice.
3. 3CalibrationCore20 minThe auction multiplies pCTR by a bid, then divides by pCTR again to set the price. An error shared by every click bidder mostly cancels, but it breaks brand bids, reserve prices and automated bidding, and an error on one slice moves money between advertisers. Measure calibration per slice and fix it with a small layer refitted every hour.
4. 4Delayed feedbackCore18 minConversions arrive hours or days after the click, so a recent click without one is not yet a negative. Treating it as one makes pCVR too low and underbids for every automated advertiser. Model the delay, or feed fake negatives and correct for them, and evaluate only on labels that have settled.

Work through the parts in order and you’ve covered the full design. Each part is one topic.