Search

Feature selection is a process of determining the most relevant features of a given problem in order to improve the generalization and the performance of a relevant classification or regression algorithm. This paper focuses on the exploitation of a genetic algorithm following a wrapping iterative approach used to extract an optimal feature subset of a large database containing pollutant concentration measurements. The feature subset is fed to a machine learning algorithm in order to predict the daily maximum concentration of two air pollutants. The encoding problem of the complexity of representation of the features in the genomes is tackled. Results of the experimentation on a specific dataset of an air quality forecasting problem are presented, as well as some proposed alterations on the standard genetic algorithm that guided the process to a mature convergence and gave good solutions for this problem. A modified version of the initial algorithm is presented as well, implemented for the purpose of being compared on an equal basis with other feature selection methods. Two such methods of the filtering type, CFS and ReliefF, are being compared with. The comparative results suggest that the wrapping type technique described in this paper is significantly better in the specific problem at hand, but this conclusion is limited to the machine learning algorithm that the technique uses at its core in the feature selection phase. [ABSTRACT FROM AUTHOR]