Learning Disentangled Latent Factors for Individual Treatment Effect Estimation Using Variational Generative Adversarial Nets

Qingsen Bao, Zeyong Mao, Lei Chen

Abstract: Estimating individual treatment effect (ITE) is a challenging task due to the need for individual potential outcomes to be learned from biased data and counterfactuals are inherently unobservable. Some researchers propose to use generative adversarial approaches to infer the counterfactual outcomes based on the distribution of factual outcomes. However, these methods assume that complete confounding factors are observed, and simply treat all observed variables as confounding factors, ignoring identification of possible instrumental factors and adjustment factors, which will bring large deviation to ITE estimation when facing biased data. To address these issues, we propose a novel Variational Generative Adversarial Nets for ITE estimation by designing the collaborative learning strategy with Variational AutoEncoder (VAE) and Generative Adversarial Nets (GAN). Specifically, we employ VAE to infer the latent representations of observed variables to access complete latent factors while using GAN to infer unseen counterfactual outcomes and guide VAE for disentangling these latent factors into three sets corresponding to the instrumental, confounding, and adjustment factors. Then the disentangled latent confounding factors can be used to further control data bias using an adaptive weighting scheme. Extensive experiments on real and synthetic data demonstrate learning disentangled latent factors for ITE estimation is effective, and our method has excellent performance even with high data bias.