Experimental comparisons with respect to the usage of the promising relations in EDA-based causal discovery

Abstract

A Bayesian network is a promising probabilistic model to represent causal relations between nodes (random variables). One of the major research issue in a Bayesian network is how to infer causal relations from a dataset by constructing better heuristic learning algorithms. Many kinds of approaches were so far introduced, and estimation of distribution algorithms (EDAs) are one of the promising causal discovery algorithms. However, the performance of EDAs is considerably dependent on the quality of the first population because new individuals are reproduced from the previous populations. In this paper, we introduce a new initialization method for EDAs that extracts promising candidate causal relations based on causal scores. Then, we used the promising relations to construct a better first population and to reproduce better individuals until the learning algorithm is terminated. Experimental results show that EDAs infer a more number of correct causal relations when promising relations were used in EDA based structure learning. It means that the performance of EDAs can be improved by providing better local search space, and it was the promising relations in this paper.