Authors: Yinghan Hong; Zhifeng Hao; Guizhen Mai; Han Huang
Addresses: School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China; School of Physics and Electronic Engineering, Hanshan Normal University, Chaozhou 521041, China ' School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China; School of Mathematics and Big Data, Foshan University, Foshan 528000, China ' School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China ' School of Software Engineering, South China University of Technology, Guangzhou 510006, China
Abstract: Feature selection is generally a key pre-process in artificial intelligence, machine learning and pattern recognition. Its purpose is to select a set of features that is most effective to predict the target. The existing features selection methods are generally a kind of features sorting methods according to the dependence between these features and the target variable. It is difficult for these methods to determine a certain number of features; moreover, in this study we show that some key feature is probably removed by these methods. To alleviate this problem, a causal feature selection method based on causal network is proposed. When the target variable and its candidate feature set form a causal network model, the proposed method can detect the causal features by conditional independence test based method according to extended Markov blanket. This method is able to cut out a certain number of features, and simultaneously can avoid missing any key feature. Experimental results demonstrate that the proposal outperforms the counterparts when applied to support vector regression.
Keywords: causal feature selection; causal network; conditional independence test; Markov blanket.
International Journal of Wireless and Mobile Computing, 2018 Vol.15 No.4, pp.310 - 317
Received: 18 Jun 2018
Accepted: 26 Aug 2018
Published online: 26 Dec 2018 *