Rethink Targeting: Detect 'Smart Cheating' In Online Advertising Through Causal Inference

In online advertising, one of the central questions of ad campaign assessment is whether the ad truly adds values to the advertisers. To measure the incremental effect of ads, the ratio of the success rates of the users who were and who were not exposed to ads is usually calculated to represent ad effectiveness. Many existing campaigns simply target the users with high predicted success (e.g. purchases, searches) rate, which often neglect the fact that even without ad exposure, the targeted group of users might still perform the success actions, and hence show higher ratio than the true ad effectiveness. We call such phenomena 'smart cheating'. Failure to discount smart cheating when assessing ad campaigns may favor the targeting plan that cheats hard, but such targeting does not lead to the maximal incremental success actions and results in wasted budget. In this paper we define and quantify smart cheating with a smart cheating ratio (SCR) through causal inference. We apply our approach to multiple real ad campaigns, and find that smart cheating exists extensively and can be rather severe in current advertising industry.