Centroid Selection in Kernel Extreme Learning Machine using K-means

Abstract

Kernel Extreme Learning Machine (KELM) is used for classification, regression, clustering and feature selection with the help of kernel functions. Conventional KELM uses all training instances as centroids for classification problem while reduced KELM uses randomly choosen training instances as centroids. Furthermore, reduced KELM is used for reducing the computational complexity of conventional KELM. To further improve the computational complexity of KELM, K-means clustering algorithm for centroid selection in KELM is proposed in this paper. In this proposed approach, number of centroids are selected as 1/10 or 5/10 of the total number of training instances and then centroids are computed by using K-means algorithm. Experiments have been carried out by using 15 data sets to illustrate the effectiveness of the proposed method. The results obtained show the reduction in computational time and increment in G-mean which verify the proposed method as an efficient approach in comparison to earlier works.