Abstract:

Support Vector Data Description (SVDD) is a typical supervised one-class classifier. To apply SVDD to unsupervised multi-class settings, a new Artificial Immune Kernel Clustering based SVDD (AIKC-SVDD) is proposed. It uses memory antibodies generated with artificial immune kernel clustering method as target data, and executes SVDD for multi-class classification. On one hand, AIKCSVDD organically combines the advantages of kernel clustering on handling unclear boundary of classification and the capacity of artificial immune network clustering on global convergence without the requirement of extra prior knowledge. On the other hand, AIKCSVDD uses memory antibodies as target data instead of original data that can better reflect the global distribution of the original data. Compared with multi-SVDD based on K-means clustering, AIKCSVDD does not need pre-determined classification amount. Experiments on artificial dataset and UCI dataset show that, under the condition of appointing classification amount for multi-SVDD, AIKCSVDD still performs better classification than multi-SVDD.

Key words: artificial intelligence, support vector domain description, artificial immune network, kernel clustering