Learning Causal Structures Based on Markov Equivalence Class

Abstract

Because causal learning from observational data cannot avoid the inherent indistinguishability for causal structures that have the same Markov properties, this paper discusses causal structure learning within a Markov equivalence class. We present that the additional causal information about a given variable and its adjacent variables, such as knowledge from experts or data from randomization experiments, can refine the Markov equivalence class into some smaller constrained equivalent subclasses, and each of which can be represented by a chain graph. Those sequential characterizations of subclasses provide an approach for learning causal structures. According to the approach, an iterative partition of the equivalent class can be made with data from randomization experiments until the exact causal structure is identified.