Problem
eBay sells promoted auction listings through Promoted Listings Express (PLX): a seller pays one flat upfront fee for more views on the auction. The listings appear in a few fixed slots of the Similar Sponsored Item block, shown next to the item being viewed (the seed item). Sellers need their listings to reach a view target before the auction ends, and buyers need slots filled with items relevant to the seed. Ranking by the seller target alone gives irrelevant recommendations, and ranking by relevance alone leaves new listings without views.
Illustrative example: a new auction listing has no click history, so a click-based score never shows it, yet its seller has paid for visibility and the auction is ending.
Solution in brief
eBay went through three rankers. A weighted random draw by urgency delivers views. A hand-weighted sum of urgency and relevance improves engagement. The current ranker draws a random number per request: with probability it uses the urgency score, otherwise a logistic regression click model, so each objective is optimized separately.
Questions this case answers
- How to rank sponsored items in a few fixed slots so that a delivery commitment to the item owner and the relevance for the viewer are both met
- How to blend a rule-based delivery objective with a learned relevance model when new items have no behavioral data
- How to choose which promoted listings to show in fixed recommendation slots when sellers need guaranteed views and buyers need relevant items?
- How to combine a seller-side delivery target and a buyer-side relevance score in one ranking?
Tools and methods
- Ratio ranking — weighted random selection by remaining views divided by remaining time, for view delivery
- Heuristic score — manual weighted combination of the seller score and a relevance score
- Epsilon-greedy mixing — random switch between the click model and the urgency score
- Logistic regression — click-probability model on more than 30 features
Google or a one-time email code. No password.