Feature Selection Algorithm Based on Discernibility Matrix and Fruit Fly Optimization

Abstract

AbstractFeature selection methods that combine swarm intelligence optimization algorithms with rough sets have been widely used. Normally, this type of method can quickly search for feature subsets, effectively reducing the computational workload; However, in some decision tables of this type of method, there were also problems such as the searched feature subset containing redundant features, which led to a decrease in accuracy. Aiming at such problems in the current feature selection algorithms based on rough sets and fruit fly optimization, improvements have been made from two aspects. Firstly, a method for measuring the importance of discernibility matrix attributes was proposed and used as a heuristic factor to constructed a new fitness function; Secondly, guided by the fitness function, a fruit fly optimized feature selection method was designed, which can avoid redundant attributes in the feature subset. Examples and experimental results on the UCI datasets show that the new feature selection method can effectively search for the smallest feature subset and improve the stability.KeywordsDiscernibility matrixFruit fly optimization algorithmAttribute importanceFeature subsetFitness function