Tournament searching method to feature selection problem

Abstract

A new search method to the feature selection problem - the tournament searching - is proposed and compared with other popular feature selection methods. The tournament feature selection method is a simple stochastic searching method with only one parameter controlling the global-local searching properties of the algorithm. It is less complicated and easier to use than other stochastic methods, e.g. the simulated annealing or genetic algorithm. The algorithm was tested on several tasks of the feature selection in the supervised learning. For comparison the simulated annealing, genetic algorithm, random search and two deterministic methods were tested as well. The experiments showed the best results for the tournament feature selection method in relation to other tested methods. © 2010 Springer-Verlag.