Global Optimization for Advertisement Selection in Sponsored Search

Abstract

Advertisement (ad) selection plays an important role in sponsored search, since it is an upstream component and will heavily influence the effectiveness of the subsequent auction mechanism. However, most existing ad selection methods regard ad selection as a relatively independent module, and only consider the literal or semantic matching between queries and keywords during the ad selection process. In this paper, we argue that this approach is not globally optimal. Our proposal is to formulate ad selection as such an optimization problem that the selected ads can work together with downstream components (e.g., the auction mechanism) to achieve the maximization of user clicks, advertiser social welfare, and search engine revenue (we call the combination of these objective functions as the marketplace objective for ease of reference). To this end, we 1) extract a bunch of features to represent each pair of query and keyword, and 2) train a machine learning model that maps the features to a binary variable indicating whether the keyword is selected or not, by maximizing the aforementioned marketplace objective. This formalization seems quite natural; however, it is technically difficult because the marketplace objective is non-convex, discontinuous, and indifferentiable regarding the model parameter due to the ranking and second-price rules in the auction mechanism. To tackle the challenge, we propose a probabilistic approximation of the marketplace objective, which is smooth and can be effectively optimized by conventional optimization techniques. We test the ad selection model learned with our proposed method using the sponsored search log from a commercial search engine. The experimental results show that our method can significantly outperform several ad selection algorithms on all the metrics under investigation.