Conditional optimization in uplift modeling

Abstract

The article is focused on a conditional optimization problem for uplift models with two given target metrics. This problem arises if we want to simultaneously maximize two metrics, for example, the customer happiness and the net profit. We present a method which maximizes the average value of one metric while the average value of another metric is fixed. The difficulty of conditional optimization is that we need to estimate the average metric value for a policy proposed by the uplift model. We cannot use the predictions of the uplift model for this estimation. We present an effective algorithm that estimates the average metric value for an arbitrary policy based on the uplift model