Maximum Relevancy Minimum Redundancy Based Feature Subset Selection using Ant Colony Optimization

In recent years dimensionality reduction of data has gained a lot of interest from the machine learning community, partly due to the huge amount of data available for processing. Classical machine learning algorithms were designed to work with limited amount of data where more emphasis was given to the learning methodology of such algorithms e.g. learning crisp rules from a dataset. Recent explosion of data, proved detrimental to the accuracy of classification algorithms due to lowering storage costs and inexpensive processing power. Therefore, feature selection which is a key technique in dimensionality reduction has become an important frontier in machine learning research. In this paper, we propose a novel filter based feature selection method. Proposed method is based on Ant Colony Optimization (ACO), and maximum Relevance and Minimum Redundancy (mRMR) for efficient subset evaluation. Although wrapper methods frequently use ACO for feature subset generation but ACO is not thoroughly studied in the development of filter methods. Proposed method takes both feature relevancy and feature redundancy in account. Our research ensures selection of features which are highly relevant with the target concept, weakly redundantly with each other and useful predictor for classification algorithms. We have performed an extensive experimentation over eleven publicly available datasets and three popular machine learning classifiers. Experimental results of comparisons show that proposed method achieves higher classification accuracy and employ reduced number of features.