A methodology to carry out voting classification tasks using a particle swarm optimization-based neuro-fuzzy competitive learning network

The problem being investigated in this paper concerns the generation of an optimal ensemble (i.e. subset) of classifiers (picked up from set of classifiers applied on a specific classification task) that maximizes the classification performance of the voting ensemble method. The design of an algorithmic framework to meet the above goal would benefit the voting process as far as its complexity is concerned. The methodology employed here treats the classifiers as objects represented by binary vectors, and quantifies the dissimilarities between pairs of classifiers. Then, a multidimensional scaling approach is put in place to transform the classifiers into points in a low-dimensional Euclidean space. The set of the resulting points is processed by a neuro-fuzzy competitive learning network trained by a hybrid procedure that combines the merits of fuzzy clustering and particle swarm optimization. The network’s outcome is a set of homogenous groups of classifiers. To this end, the optimal ensemble is obtained by selecting from each group the best classifier. The method was successfully applied to a number of simulation experiments that involved many data sets and classifiers. Comparative analysis with other relative algorithms took place. The results were very promising as the proposed method appeared to be significantly superior in all of the experiments.