An Improved Optimal Pairwise Coupling Classifier 1 Roger Xu 1 , Tao Qian 1 , Chiman Kwan 1 1 Intelligent Automation, Inc., 15400 Calhoun Drive, Suite 400, Rockville, MD 20855 {hgxu, tqian, ckwan}@i-a-i.com Bruce Linnell and Tim Griffin NASA KSC Abstract. The Optimal Pairwise Coupling (O-PWC) classifier was proposed and used for data classification because of its excellent classification performance [2]. A key step in the O-PWC algorithm is to calculate a number of posterior probabilities, which was achieved using an iterative procedure in [1],[2]. In this paper, we will present an analytical solution to the problem of finding the posterior probabilities. As a result, the computational efficiency of the O-PWC algorithm will be significantly improved. One numerical example will be given to show the improved computational efficiency of the improved O-PWC classification algorithm. We will also present the classification results using O-PWC and compare its performance with 3 other conventional classification techniques. 1 Introduction For a K-class classification problem, the one-against-one method is being widely used. In this method, K(K-1)/2 two-class classifiers are constructed. Each of them is utilized to discriminate two of the K-classes. MaxVoting (MV) and Pairwise Coupling (PWC)[1] are two common ways to combine the results of these binary classifiers. MV considers the output of each classifier as a binary decision and selects the class that wins the most votes. In PWC, however, each of the binary classifier outputs a posterior probability, called pairwise probability, for a given test sample. PWC combines these pairwise probabilities into K posterior probabilities. PWC selects the class that has the largest posterior probability. In many applications PWC performs better than MV. However, the original PWC has some drawbacks. When a test sample is classified by one of the K(K-1)/2 classifiers and it doesn’t belong to either of the two involved classes of this classifier, the probabilistic measures of the test sample against the two classes become meaningless and therefore may damage the final fused output of PWC. One new approach called optimal PWC (O-PWC)[2] was proposed to tackle this problem. The key idea of O-PWC is as follows. For the same K-class classification problem, an array of K O-PWC classifiers is constructed, each of which 1 Research supported by NASA Kennedy Space Center under contract NAS10-03022