Exploitation and comparison of computational intelligence techniques on the feature selection problem

Abstract

Feature selection aims to gain a minimal feature subset in a problem domain while conserving the accuracy of the original data. Feature selection is a process for making more efficient data analysis by selecting more relevant features for the related problem solving. Feature selection increases prediction probability of algorithms by decreasing the dimensionality, eliminating irrelevant features. In this study, four computational intelligence techniques are implemented and compared on the well-known data instances taken from UCI. This is the first time that Migrating Birds Optimization (MBO) is used for the feature selection problem. Specifically, the exploited algorithms are (i) MBO, which is a recently proposed but successful technique, (ii) particle swarm optimization, which has originated from the simulation of behavior of biological organisms, (iii) simulated annealing, which is a well-known and frequently used as a benchmark algorithm and (iv) differential evolution. In our filter-based approach, we also implemented the inconsistency based subset evaluator to evaluate the performance of a given feature subset. Performance comparison is done with k-nearest neighbor, as the classifier where all features are used in the benchmark. Results show that the MBO algorithm presents the best performance in terms of number of winning cases.