A Novel Multi-classification Method Based on One-against-One Relevance Vector Machine

Abstract

Relevance Vector Machine (RVM) has excellent classification performance for nonlinear, high dimensional and small data sets. Traditional One-against-One (OAO) is applied the most widely by it in multiclassification. However, it uses many classifiers which results in slow classification. In order to improve the classification speed, a new method based on OAO is proposed. The method finds the class corresponding to the test sample from a narrower class ranges through circular computation. At the classifying prediction stage, all the possible classes are put as a circle; each adjacent class is classified; the classes with most votes are the possible classes of next iteration. After repeating these processes, the last remaining one is the predicted class. Experiments show that the proposed method can ensure classification accuracy while enhancing the speed effectively.