A Hybrid Feature Selection Algorithm Based on Large Neighborhood Search

Abstract

Feature selection aims at choosing a small number of relevant features in a data set to achieve similar or even better classification accuracy than using all features. This paper presents the first study on Large Neighborhood Search (LNS) algorithm for the feature selection problem. We propose a novel hybrid Wrapper and Filter feature selection method using LNS algorithm (WFLNS). In LNS, an initial solution is gradually improved by alternately destroying and repairing the solution. We introduce the idea of using filter ranking method in the process of destroying and repairing to accelerate the search in identifying the core feature subsets. Particularly, WFLNS either adds or removes features from a candidate solution based on the correlation based feature ranking method. The proposed algorithm has been tested on twelve benchmark data sets and the results have been compared with ten most recent wrapper methods where WFLNS outperforms other methods in several the data sets.