Sponsored Search Ad Ranking

Sponsored search ad ranking is the online-learning problem of choosing which advertisement to display for a search query, learning from click feedback while balancing relevance, revenue, and exploration.

Key points

• Lu, Pál, and Pál use sponsored web search as the motivating application for Lipschitz Contextual Bandits: a query arrives, the system displays an ad, and the click outcome provides bandit feedback ^[src-020].
• The unknown payoff is click-through rate for a query-ad pair. A Bayes-optimal policy would display the highest-CTR ad for each query, but the CTR function must be learned over time ^[src-020].
• The exploration-exploitation trade-off is direct: a search engine should exploit ads that appear relevant while still exploring ads that might be better for similar queries ^[src-020].
• The Lipschitz assumption gives the system a way to generalise: similar queries and similar ads should have nearby click-through rates, allowing data from one region of the metric space to inform nearby decisions ^[src-020].

Related entities

_(none yet)_

Related concepts

• Contextual Bandits
• Lipschitz Contextual Bandits
• Query-Ad-Clustering
• Exploration-Exploitation Trade-off

Source references

• ^[src-020] Tyler Lu, David Pál, Martin Pál — “Contextual Multi-Armed Bandits” (AISTATS 2010)