A Fast Markov blanket discovery algorithm
Xiaofeng Zhu, Youlong Yang · 2014
Learning Markov blanket MB plays an important role in feature selection for classification, causal discovery, and Bayesian Networks learning. In this paper, an efficient and effective algorithm, called Fast Iterative Parent-Child based search of MB (FIPC-MB) is proposed to learn the MB of the target variable T. Foremost, we combined the IPC-MB algorithm with mutual information knowledge to initialize candidate parents and children (Cand_PC) of the the target node. Furthermore, we changed the sequence of variables belonging to Cand_PC(T). Finally, we employed the property of mutual information between two variables to select condition set instead of randomly choosing it from Cand_PC(T) for every conditionnal independence test(CI test). These operations drastically improve the efficiency of searching for condition set and decrease the number of CI tests. In addition, simulation experiments demonstrate that the FIPC-MB algorithm outperforms the state-of-the-art algorithm, IPC-MB, in terms of running efficiency and accuracy of performance.