Ensemble Method of Feature Selection Using Filter and Wrapper Techniques with Evolutionary Learning

The improvement in data collection and mining methods has expanded the range of dimensionality or features in the data, which brings about an obstacle to many existing feature selection methodologies. This paper brings forth a fresh feature selection methodology rooted in particle swarm optimization (PSO) as wrapper method and an ensemble method to merge the results of the different filter techniques (chi-square, F-regression, and mutual information) to find an optimal feature set that covers most of the key variables of the dataset. The local search is executed on the global best and makes use of a filter-based method, which then intends to take the advantage of the filter and wrapper methods. Our results exhibit that the proposed methodology can be successfully used to select fewer features and, at the same time, increase the classification efficiency over using all features. The proposed methodology also shows how well an evolutionary learning algorithm like the particle swarm optimizer can be used for search optimization of optimum features in the dataset.