A novel feature selection using Markov blanket representative set and Particle Swarm Optimization algorithm

Feature selection based on Markov blankets and evolutionary algorithms is a key preprocessing technology of machine learning and data processing. However, in many practical applications, when a data set does not satisfy the condition of fidelity, it may contain multiple Markov blankets of a class attribute. In this paper, a hybrid feature selection algorithm based on Markov blanket representative set via Particle Swarm Optimization is proposed to solve the problem of data classification which does not meet the condition of fidelity. The algorithm uses the maximum information coefficient to determine the correlation and redundancy between features and class attributes and among the features. It redefines the approximate Markov blanket representative set of the class attribute C which does not consider whether the data set satisfies the condition of fidelity. Then obtains the suboptimal feature subset of the original feature set. At the same time, the fitness function which combines the classification prediction ability of the feature subset and the number of selected features is introduced. On the reduced feature set, Particle Swarm Optimization algorithm is used to search for a better feature subset. A series of experiments on benchmark datasets show that the hybrid feature selection algorithm based on Markov blanket and PSO outperforms the Markov blanket-based feature selection selectors and other well-established feature selection methods.