A wrapper feature selection approach using Markov blankets

Atif Hassan, Jiaul Hoque Paik, Swanand R. Khare, Asif Hassan Syed

Abstract: Highlights•MB discovery algorithms are traditionally recognized as filter methods requiring different CI tests for different datatypes and tasks.•We propose a universal Markov blanket based novel wrapper feature selection algorithm.•A novel wrapper-based non-parametric conditional independence (CI) test is used to accurately test for CI between input variables.•An optional MB aggregation step is also proposed to potentially find the best MB set under non-faithful conditions.