Iterative Causal Discovery with Active Interventions

Anonymous

Keywords: causal discovery, causal inference, multi-entity prediction, intervention

TL;DR: We propose a causal discovery algorithm that can iteratively and actively gather interventional data to improve both causal discovery and downstream prediction performance.

Abstract: Multi-entity causal discovery is a fundamental problem in machine learning. Understanding the underlying causal relations is important for counterfactual reasoning and robustness. However, the causal structure is only identifiable up to a Markov Equivalence Class with observational data. In real life, it is usually hard or even unethical to gather interventional data. Fortunately, the presence of simulators allows for the production of real or simulated intervention data that can help in identifying causal graphs. We propose a causal discovery algorithm that can iteratively and actively gather intervention data to improve the prediction of causal graphs. We demonstrate on several datasets that iterative interventional data augmentation improves both causal discovery and dynamics prediction performance.