Factoring past exposure in display advertising targeting.

Online advertising is becoming more and more performance oriented where the decision to show an advertisement to a user is made based on the user's propensity to respond to the ad in a positive manner, (e.g., purchasing a product, subscribing to an email list). The user response depends on how well the ad campaign matches to the user's interest, as well as the amount of user's past exposure to the campaign - a factor shown to be impactful in controlled experimental studies. Past exposure builds brand-awareness and familiarity with the user, which in turn leads to a higher propensity of the user to buy/convert on the ad impression. In this paper we propose a model of the user response to an ad campaign as a function of both the interest match and the past exposure, where the interest match is estimated using historical search/browse activities of the user. The goal of this paper is two-fold. First, we demonstrate the role played by the user interest and the past exposure in modeling user response by jointly estimating the parameters of these factors. We test this response model over hundreds of real ad campaigns. Second, we use the findings from this joint model to identify more relevant target users for ad campaigns. In particular, we show that on real advertising data this model combines past exposure together with the user profile to identify better target users over the conventional targeting models.