Prediction and Exploration Models for Responsive Search Ads

Search-ads are an important product in computational advertising. In order to achieve high performance in search-ads, it is critical to improve ad-copy relevance and click-through rates. Traditionally, the onus has been on advertisers to come up with an ad copy and set up appropriate bids that can achieve these challenging objectives. Thanks to the emergence of AI and machine-learning technologies, we can simplify both the bidding and ad copy-generation processes. Response search ads (RSA) is an emerging product in this direction that tackles the ad-copy  generation piece. RSA provides advertisers with sufficient control over the adcopies that can be shown for their products by giving them the option of providing components (title and descriptions assets) of the ad copy, as opposed to the ad copy itself, and tackling the assembly of these components dynamically at runtime. This allows us to take into consideration query and user contextual information and product better ads. Even with a small number of assets, there are many potential asset combinations. In this work, we propose a large-scale logistic-regression model, combined with a sequential contextual-bandits framework that allows us to assemble high-performing adcopies at runtime. We have productized our algorithms as part of Bing Ads. The shipped models have achieved 6% adoption based on servable customer count and have already shown increasing daily revenue as well as promising marketplace KPI trends (all up RPM, CY and IY) in just a few months’ time. In this paper, we will discuss the modeling details, measurement techniques as well as online flight evaluation with real search traffic from bing.com.