Modeling labels for conversion value prediction

In performance based digital advertising, one of the key technical tools is to predict the expected value of post ad click purchases (a.k.a. conversions). Some of the salient aspects of this problem such as having a non-binary label and advertisers reporting the label in different scales make it a much harder problem than predicting probability of a click. In this paper we ask what is a good way to model the label and extract as much information as possible from the features. We investigate three main issues that arise from advertiser reported labels and come up with new techniques to address them. The first issue is that the label scale can affect how the model capacity is devoted to different advertisers. The second issue is how outlier labels can cause over-fitting. Finally, we also show that the distribution of the label contains vital information and the we train our models to use them and not just rely on the mean.