Multihead Causal Distilling Weighting Is All You Need for Uplift Modeling

Abstract

Uplift modeling is a branch of predictive modeling technology, which is usually used to analyze marketing, advertising and product personalization experiments. In this kind of application scenario, we usually do a large number of random experiments to assist decision-making. And thus there is a great need for us to control the number of features as well as tell the most important features in order for the next steps. What's more, in this kind of scenario, the interpretability of features also becomes very important, which means that we should not only pursue the accuracy of prediction, but also minimize the difficulty of controlling the features. However, most existing studies only focus on the accuracy of model prediction, but ignore the cost of controlling and observing too many variables as well as the interpretability of features in practical application. In order to better solve this problem, we introduce a multi head weight calculation method based on causal inference. Instead of selecting features based on the result of machine learning, we manage to select features from the source with causal inference, a total different method from traditional machine learning methods. It can be viewed as a method to solve the problem of overfitting. In our experiment, we use the data from different industries and use different numbers of selected features to evaluate the effectiveness of the proposed feature selection method. The results show that our algorithm significantly improves the performance compared with the feature selection method in standard machine learning theory.