Problem
trivago is a hotel metasearch. For each search it must pick about 25 accommodations out of thousands, and the ranking model favors ones with a good history. Accommodations that users never saw far outnumber the ones they see (cold start), so promising inventory stays hidden. Users miss better options, advertisers get little conversion data for bidding, and the marketplace gets less healthy over time. Exploring costs short-term user and revenue metrics, so the amount of it must be controllable.
Example from the article: a ranking feature held the mean of a beta-binomial posterior of past performance. The author calls this too greedy: it purely exploits well-performing accommodations.
Solution in brief
Replace the posterior mean in the ranking feature with an optimistic estimate: mean plus a fraction of the standard deviation, scaled by a parameter . Items with few impressions have a wide posterior and get a boost. A model trained on well-known inventory then scores unseen items, so exploration targets the most promising ones.
Questions this case answers
- How to add controlled exploration of items with no history to a ranking model without hurting topline metrics
- How to pick the most promising items among a large pool that has never been shown
- How to show never-impressed items in a ranking without losing short-term revenue?
- How to control the amount of exploration in a ranking model with a single parameter?
Tools and methods
- Multi-armed bandit view — exploitation of known good items against exploration of unseen ones.
- Beta-binomial posterior — the historical-performance feature of the ranking model.
- Optimistic estimate — boosts low-impression items; sets exploration strength.
- Item-feature model — predicts performance of unseen items.
- A/B tests and Pareto front — compare variants by extent, cost and quality of exploration.
Google or a one-time email code. No password.