The Support Vector Clustering Neural Network Used For Pattern Classification

Abstract

In this paper, we present a novel neural network called Support Vector Clustering Neural Network (SVCNN). The theoretical foundation of the network is based on Support Vector Clustering (SVC). The training method of SVC is in a batch learning mode which has a drawback that the solution of the Lagrange multipliers is difficult to find when the training data become large. To overcome this drawback, we propose a mechanism, namely Support Vector Identifying Support Vector (SVISV), and develop its associated algorithm whose training method adopts an incremental learning mode. In each step, only one or few new data attend the SVC calculation. Then the data that just act as support vectors will remain to continuously attend the calculation in the next step. Until all the training data have been clustered, the learning process is terminated. The clustering result is further to construct a SVCNN. Using the mechanism SVISV, the SVCNN can determine whether an unknown data belongs to one cluster that has been already built in the network. The simulation outcomes reveal that our SVISV algorithm is slightly faster than the traditional SVC while support vectors are the minorities in a data set.