Using boosted trees for click-through rate prediction for sponsored search

Abstract

We describe a new approach to solving the click-through rate (CTR) prediction problem in sponsored search by means of MatrixNet, the proprietary implementation of boosted trees. This problem is of special importance for the search engine, because choosing the ads to display substantially depends on the predicted CTR and greatly affects the revenue of the search engine and user experience. We discuss different issues such as evaluating and tuning MatrixNet algorithm, feature importance, performance, accuracy and training data set size. Finally, we compare MatrixNet with several other methods and present experimental results from the production system.