Conditional Causal Discovery

Cixuan Zhang, Benjie Wang

Keywords: bayesian causal discovery

Abstract: Multivariate causal discovery is the process of inferring a causal graphical model given data generated from that model. Many methods have been developed both for deriving point estimates as well as quantifying uncertainty over causal graphs. However, integrating complex knowledge or "what-if" scenarios, such as an observed causal effect value, into the discovery process remains a significant challenge. In this paper, we follow the Bayesian approach to causal discovery, and propose to cast such problems as conditional inference. The key computational challenge is that such knowledge may be very unlikely, making approaches based on naive sampling infeasible: that is, there may be no sample which satisfies the condition. To overcome this, we leverage techniques from the rare event estimation literature. Our empirical results on synthetic data illustrate the efficacy of our approach where baselines fail to accurate capture the conditional distribution, and we illustrate its application to the real-world Sachs protein dataset.

Submission Number: 17