Counterfactual inference and counterfactual data generation through latent disentanglement

Abstract

In causal graphs, inferring exogenous variables is required to identify the counterfactual effect from observations. Given that the exogenous variable is often latent in Bayesian network, the modelers have to assume the structure of exogenous variables in a causal graph, and its corresponding variational autoencoder. A frequent assumption is defining a single latent variable to absorb the entire exogenous uncertainty, but we claim that such structure cannot avoid the dilemma of 1) the biased sampling in the decoder learning and 2) the information loss to regularize the decoders of interventions. Our model resolves this dilemma by disentangling the exogenous uncertainty into two latent variables of 1) independent to interventions and 2) correlated to interventions without causality. Particularly, our disentangling approach will preserve the latent variable correlated to interventions in generating counterfactual cases. We show that our method estimates total effect and counterfactual effect without a complete causal graph. Our first application is generating counterfactual fair data for a fairness task. Our second application is generating the counterfactual image, which do not naturally occur in the dataset.