Learning Optimal Interventions

Our goal is to identify beneficial interventions from observational data. We consider interventions that are narrowly focused (impacting few features) and may be tailored to each individual or globally enacted over a population. If the underlying relationship obeys an invariance condition, our approach can identify beneficial interventions directly from observational data, side-stepping causal inference. We provide theoretical guarantees for predicted gains when the relationship is governed by a Gaussian Process, even in settings involving unintentional downstream effects. Empirically, our approach outperforms causal inference techniques (even when our model is misspecified) and is able to discover good interventions in two practical applications: gene perturbation and writing improvement.