Similarity Models for Ad Relevance Measures

Measuring the relevance between the query and paid search ads is an important problem to ensure the overall positive search experience. In this paper, we study experimentally the effectiveness of various document similarity models based solely on the content analysis of the query and ad landing page. Our approaches focus on two different aspects that aim to improving the document representation: one produces a better term-weighting function and the other projects the raw term-vectors to the concept space. Both models are discriminatively trained and significantly outperform the baseline approach. When used for filtering irrelevant ads, combining these two models gives the most gain, where the uncaught bad ads rate has reduced 28.5% when the false-positive rate is 0.1.