25
4
v1v2 (latest)

Learning and Sampling of Atomic Interventions from Observations

Abstract

We study the problem of efficiently estimating the effect of an intervention on a single variable (atomic interventions) using observational samples in a causal Bayesian network. Our goal is to give algorithms that are efficient in both time and sample complexity in a non-parametric setting. Tian and Pearl (AAAI `02) have exactly characterized the class of causal graphs for which causal effects of atomic interventions can be identified from observational data. We make their result quantitative. Suppose P is a causal model on a set V\vec{V} of n observable variables with respect to a given causal graph G with observable distribution PP. Let PxP_x denote the interventional distribution over the observables with respect to an intervention of a designated variable X with x. Assuming that GG has bounded in-degree, bounded c-components (kk), and that the observational distribution is identifiable and satisfies certain strong positivity condition, we give an algorithm that takes m=O~(nϵ2)m=\tilde{O}(n\epsilon^{-2}) samples from PP and O(mn)O(mn) time, and outputs with high probability a description of a distribution P^\hat{P} such that dTV(Px,P^)ϵd_{\mathrm{TV}}(P_x, \hat{P}) \leq \epsilon, and: 1. [Evaluation] the description can return in O(n)O(n) time the probability P^(v)\hat{P}(\vec{v}) for any assignment v\vec{v} to V\vec{V} 2. [Generation] the description can return an iid sample from P^\hat{P} in O(n)O(n) time. We also show lower bounds for the sample complexity showing that our sample complexity has an optimal dependence on the parameters nn and ϵ\epsilon, as well as if k=1k=1 on the strong positivity parameter.

View on arXiv
Comments on this paper