Using classifier fusion techniques for protein secondary structure prediction

Classifier fusion techniques are gaining more popularity for their capability of improving the accuracy achieved by individual classifiers. A common approach is to combine the classifiers' outcome using simple methods, such as majority voting. In this paper, we build a meta-classifier by fusing some already well-known classifiers for protein structure prediction. Each individual classifier outputs a unique structure for every input residue. We have used the confusion matrix of each protein secondary structure classifier, which is representative of classifiers' expertness, as a general reusable pattern for converting its simple class-label assignment to class-preference score. The results obtained using several classifier fusion operators have been compared, on some standard datasets from the EVA server, with simple majority voting and with the results provided by the individual classifiers. The comparative analysis showed that the Choquet fuzzy integral operator had the highest improvement with respect to accuracy, multi-class sensitivity and specificity criteria over both the best performing individual classifier and the other fusion operators, while all of the classifier fusion techniques yielded some improvements too.