Classifier combination through clustering in the output spaces

Abstract

This paper proposes the use of information about the distribution of the classifier outputs in their output spaces during combination. Two different methods based on the clustering in the output spaces are developed. In the first approach, taking into account the distribution of the output vectors in these clusters, the local reliability of each individual classifier is quantified and used for weighting the classifier outputs during combination. In the second method, the classifier outputs are replaced by the centroids of the nearest clusters during combination. Experimental results have shown that both of the proposed approaches provide more than 3% improvement in the correct classification rate.