Kernel evolutionary algorithm for clustering

Abstract

In this paper, we propose a novel clustering algorithm named KECA based on kernel function and evolutionary optimization. As we know, Euclidean distance based similarity metrics can help clustering algorithms handle datasets with compact super-sphere distributions perfectly, but it is undesirable for the complex structural or irregular shaped datesets. Proper mapping function can map the data in original space to high-dimensional feature space, which exposes more features and sheds light on complex structural datasets. However, clustering in feature space is time-consuming and often suffers from curse of dimensionality. Fortunately, we can cluster the mapped data in feature space which performs nonlinearly in original space with the help of kernel function in our proposed KECA. What’s more, evolutionary algorithm is used in KECA to avoid local optimal. Experimental results on artificial as well as UCI datasets show the effectiveness and robustness of the proposed KECA in compare with the genetic algorithm-based clustering and the K-means clustering.