Online Learning of a Ranking Formula for Revenue and Advertiser ROI Optimization

A standard model for sponsored search comprises of ranking ads by their expected revenue, that is, the product of their bid price and estimated click-through rate (CTR). In this work, we introduce two complementary use cases for ranking ads at an online classifieds site and aim to optimize a ranking formula which extends the traditional one. First, we address the task of ranking ads on the search results page for revenue optimization. While most works address this challenge by improving CTR estimation, we consider the effectiveness of the CTR estimation as a given and presume that if CTR estimation is somewhat ineffective, it can be compensated by applying a larger weight to the bid factor. Second, we aim to improve advertiser return on investment (ROI) while keeping a similar level of revenues for ads ranking on the home page feed. To this end, we introduce into the standard ranking formula a factor that favors ads with higher click-out rate and serves as an effective tie-breaker in cases of two competing ads with relatively similar revenue expectations. To optimize the ranking formula, for each case, we propose an online learning procedure in a multi-armed bandit setting. Empirical evaluation attests to the merits of this approach compared to the existing ranking in production, which is based on the traditional formula, and validates our reasoning, first, regarding the relationship between CTR estimation effectiveness and the learned weights, and second, on the contribution of the click-out factor to increase in advertiser ROI.