Causal normalizing flows: from theory to practice
Keywords: causality, causal inference, normalizing flows, identifiability, interventions, counterfactuals
TL;DR: Armed with identifiability results, we demonstraste how to use normalizing flows to capture a causal model and perform causal inference with it.
Abstract: In this work, we bridge the gap between normalizing flows and causal inference. First, we leverage recent results on non-linear ICA to show that causal models are identifiable from observational data given a causal ordering, and thus can be recovered using autoregressive normalizing flows (NFs). Second, we propose a simple design for causal normalizing flows to ease learning and capture the underlying causal data-generating process. Third, we describe how to implement the do-operator in causal NFs, and thus, how to answer interventional and counterfactual questions. Finally, we empirically validate our proposed design by comparing causal NFs to other approaches for approximating causal models, and demonstrate that causal NFs can be used in real-world problems—where mixed discrete-continuous data and partial knowledge on the causal graph is the norm.
Submission Number: 19
Loading