A new approach to learn the projection of latent Causal Bayesian Networks

Due to limitations of the total cost of randomized controlled experiments and latent variables exist, it is difficult to learn a graph for indicating the true causal relations in original graph. This paper presents a new approach which contains two stages to learn a projection of Causal Bayesian Networks with unobserved variables. The first stage is to learn a skeleton of projection of Causal Bayesian Networks by using a new algorithm LSofP. It reduces computation amount of and improves reliability of conditional independence tests compared with other existing algorithms. At the same time, algorithm LSofP also does not introduce spurious links and directed edges in graph returned represent the true causal relations. So the structure learned is more accurate. In order to orient as more edges as possible, prior knowledge and an optimal experiment design are incorporated in the second stage by using algorithm IPKOED. Theoretical results show that the new approach is correct and efficient, and a better representation of Causal Bayesian Networks returned. Simulation results illustrate its advantages over the existing algorithms.