Missing Click History in Sponsored Search: A Generative Modeling Solution

A fundamental problem in sponsored search advertising is the estimation of probability of click for ads displayed in response to search queries. The historical click-through rate (CTR) is one of the most important predictors of the click, and extracted at multiple resolutions of the query-ad hierarchy. However, the new ads do not have any click history, and even the existing ads might miss history at some resolutions due to, for example, tail queries. In addition to a loss in accuracy, the missing features introduce significant complexity in designing conditional probability of click models such as the maximum-entropy model. In this paper, we develop a generative modeling solution to handle missing features in the maximum-entropy and other conditional models. In particular, a mixture of multivariate Gaussian distributions is used to learn a representation of the CTR features. This mixture model then provides information about the missing features to the maximum-entropy model in increasing degrees of sophistication, ranging from the pointwise estimates of the missing features to multi-way interaction terms and to novel posterior features. We show the utility of this approach for sponsored click prediction using the click-view data collected from Yahoo! search engine. We find that the generative modeling approach not only improves click prediction accuracy over a state-of-the-art system, but also results in a significantly less complex system.