CAUSAL NEURAL NETWORKS FOR CONTINUOUS TREATMENT EFFECT ESTIMATION

Abstract

Causal inference have wide applications in medical decision-making, evaluating advertising, and voucher distribution. The exist of confounding effect makes it difficult to have an unbiased uplift estimation. Traditional methods focuses on the ordering of the problem. Little attention have been paid to the response performance, either on the evaluation metric, nor the modeling. In this work, an end-to-end multi-task deep neural network is proposed to capture the relations between the treatment propensity and the treatment effect, where the treatment can be continuous. The performance of the proposal is tested over large scale semi-synthetic and real-world data. The result shows that the proposal balances the estimation of response performance and individual treatment effect. The online environment implementation suggests the proposal can boost up the market scale and achieve 4.8% higher return over investment (ROI).