On Noise Abduction for Answering Counterfactual Queries: A Practical Outlook

Abstract

A crucial step in counterfactual inference is abduction - inference of the exogenous noise variables. Deep Learning approaches model an exogenous noise variable as a latent variable. Our ability to infer a latent variable comes at a computational cost as well as a statistical cost. In this paper, we show that it may not be necessary to abduct all the noise variables in a structural causal model (SCM) to answer a counterfactual query. In a fully specified causal model with no unobserved confounding, we also identify exogenous noises that must be abducted for a counterfactual query. We introduce a graphical condition for noise identification from an action consisting of an arbitrary combination of hard and soft interventions. We report experimental results on both synthetic and real-world German Credit Dataset showcasing the promise and usefulness of the proposed exogenous noise identification.