Multi Center Polyhedral Conic Classifiers for Estimating Non-Linear Decision Boundaries

摘要

Polyhedral conic classifiers are getting popular with the performance against support vector machines (SVM). In these classifiers a conic function with a vertex point is used. Vertex point is an important parameter and improves the performance when it is set to the mean of positive samples. In cases where positive data belonging to the same class are clustered in different regions, a single classifier is not enough, and more than one classifier is needed. In this study, a novel multi center polyhedral conic classifiers (MCPCC) method is developed to use only one classifier to classify positive data clustered in different centers. Experiments are performed on two synthetic datasets using proposed method. The proposed method was compared with Kernel SVM and other polyhedral conic classifiers and it was found to give impressive results.