Auto IV: Counterfactual Prediction via Automatic Instrumental Variable Decomposition

Instrumental variables (IVs), sources of treatment randomization that are conditionally independent the outcome, play an important role in causal inference with unobserved confounders. However, existing IV-based counterfactual prediction methods need well-predefined IVs, while it’s art rather than science to find valid IVs many real-world scenes. Moreover, predefined hand-made could be weak or erroneous by violating conditions IVs. These thorny facts hinder application methods. In this article, we propose a novel Automatic Variable decomposition (AutoIV) algorithm automatically generate representations serving from observed (IV candidates). Specifically, let learned IV satisfy relevance condition and exclusion outcome via mutual information maximization minimization constraints, respectively. We also learn confounder encouraging them relevant both outcome. The compete for their constraints adversarial game, which allows us get prediction. Extensive experiments demonstrate our method generates accurate