One-against-all multi-class SVM classification using reliability measures

Support vector machines (SVM) is originally designed for binary classification. To extend it to multi-class scenario, a typical conventional way is to decompose an M-class problem into a series of two-class problems, for which one-against-all is the earliest and one of the most widely used implementations. However, certain theoretical analysis reveals a drawback, i.e., the competence of each classifier is totally neglected when the results of classification from the multiple classifiers are combined for the final decision. To overcome this limitation, this paper introduces reliability measures into the multi-class framework. Two measures are designed: static reliability measure (SRM) and dynamic reliability measure (DRM). SRM works on a collective basis and yields a constant value regardless of the location of the test sample. DRM, on the other hand, accounts for the spatial variation of the classifier's performance. Based on these two reliability measures, a new decision strategy for the one-against-all method is proposed, which is tested on benchmark data sets and demonstrates its effectiveness.