A Combine-Correct-Combine Scheme for Optimizing Dissimilarity-Based Classifiers

摘要

Recently, to increase the classification accuracy of dissimilarity-based classifications (DBCs), Kim and Duin [5] proposed a method of simultaneously employing fusion strategies in representing features (representation step) as well as in designing classifiers (generalization step). In this multiple fusion strategies, however, the resulting dissimilarity matrix is sometimes an indefinite one, causing problems in using the traditional pattern recognition tools after embedding the matrix in a vector space. To overcome this problem, we study a new way, named combine-correct-combine (CCC) scheme, of additionally employing an Euclidean correction procedure between the two steps. In CCC scheme, we first combine dissimilarity matrices obtained with different measures to a new dissimilarity representation using a representation combining strategy. Next, we correct the dissimilarity matrix using a pseudo-Euclidean embedding algorithm to improve the internal consistency of the matrix. After that, we again utilize the classifier combining strategies in the refined dissimilarity matrix to achieve an improved classification for a given data set. Our experimental results for well-known benchmark databases demonstrate that the CCC mechanism works well and achieves further improved results in terms of the classification accuracy compared with the previous multiple fusion approaches. The results especially demonstrate that the highest accuracies are obtained when the refined representation is classified with the trained combiners.