Uplift Prediction with Dependent Feature Representation in Imbalanced Treatment and Control Conditions

Artem Betlei, Eustache Diemert, Massih-Reza Amini

Abstract: Uplift prediction concerns the causal impact of a treatment over individuals and it has attracted a lot of attention in the machine learning community these past years. In this paper, we consider a typical situation where the learner has access to an imbalanced treatment and control data collection affecting the performance of the existing approaches. Inspired from transfer and multi-task learning paradigms, our approach overcomes this problem by sharing the feature representation of observations. Furthermore, we provide a unified framework for the existing evaluation metrics and discuss their merits. Our experimental results, over a large-scale collection show the benefits of the proposed approaches.