A New Score Correlation Analysis Multi-class Support Vector Machine for Microarray

This paper investigates multi-class support vector machines (SVM). The main objective is to improve the classification accuracy of SVM on microarray data of multiple-category. This is achieved by introducing a new voting approach to the assignment of class label for a test observation after pairwise training of SVM classifiers. The approach investigates the correlations between "scores" — the real valued vector produced for observations by a set of binary SVM classifiers. These score vectors are then combined using a majority voting mechanism to assign the class membership for the test observations. The performance of the algorithm is evaluated on various gene expression profiles, and two typical multi-class SVM algorithms, namely the max-wins voting by Friedman and pairwise coupling by Hastie and Tibshirani, are compared with the proposed method. The experimental results on synthetics data and microarray of real life show the efficacy of the proposed method and that the new multi-class SVM is superior to max-wins and pairwise coupling in terms of the classification of multiple-labeled microarray.