Finally, five E7050 solubility υ-SVM sub-classifiers with roughly high correctness rate were selected using majority voting method. The success of majority voting depends on the number of members in the voting group. In this paper, we investigate the number of members in a majority voting group that gives the best results. A lot of experimental results indicate
that performing ICA process and selecting a set of ICs to reconstruct samples, makes correctness rate of υ-SVM sub-classifiers unstable. Thus, an appropriate number of sub-classifiers have to be trained to display all possible results. In this paper, four experiments have been carried out on 3 data bases. In Tables Tables11--3,3, minimum and maximum amounts of 25 υ-SVM sub-classifiers and also general correctness rate is demonstrated. Furthermore, Figures Figures22--44 demonstrate correctness rate
box plot respectively in 4 experiments, as x and y axis are demonstrators of the number of test samples and correctness rate of the classifier, respectively. From Figures Figures33--5,5, it is observed that if a greater number of ICs are removed, five existing amounts in box-plots related to microarray data (minimum, first quadrature, medium, third quadrature, and maximum) will decline (except in the third experiment related to lung cancer). This subject shows that correctness rate of classifier changes according to the number of used ICs to reconstruct. If a greater number of ICs are removed, general correctness rate of the classifier
proportioned to each sub-classifier will improve, apparently. Drug_discovery Similar results can be achieved in Tables Tables11--3.3. As can be seen, correctness rate related to the whole classifier is more than correctness rate related to each classifier. For example, the ensemble correctness rate for 7 IC components in Leukemia dataset is 0.9444, while the maximum and minimum correctness rates for the same IC components in this dataset are 0.9306 and 0.8472, respectively. This point is worth noticing that in case of removing more ICs, classifier performance faces problem and becomes unstable. Thus, a trade-off must be established between the number of ICs used for reconstruction and correctness rate of the classifier.