A new RBF kernel based learning method applied to multiclass dermatology diseases classification

Feature selection is a vital process in classification of medical datasets. This paper addresses feature selection in Radial Basis Function (RBF) kernel space for the classification of multiclass dermatology dataset using neural network and data mining classifiers. It has three stages in determining relevant and irrelevant features for the classification task. In stage I, the features of dermatology diseases dataset are transformed to RBF kernel space. In stage II, kernel mean of the transformed features are computed using the values obtained from multiclass improved F-Score formula. In stage III, the features greater than kernel mean are used in classification process with Support Vector Machines (SVM), Radial Basis Function Network (RBFN) and C4.5. The dermatology diseases dataset is taken from machine learning repository, University of California, Irvine. It contains 34 features, 366 instances and 6 classes. When evaluated, this new method of feature selection carried out in RBF kernel space has a peek performance and resulted 96.0% (Ten-fold cross validation) of classification accuracy for C4.5 which is higher than the results obtained in original space. The results indicate that feature selection carried out in RBF kernel space is promising than the results obtained in original space for multi class datasets.