Selective Ensemble Method Based on Spectral Clustering

Abstract

Currently, spectral clustering has attracted wide attention in clustering analysis domain. Compared with traditional partition-based clustering algorithm, spectral clustering is more adaptable to data distribution and less computational complexity for the large-scale data sets. Nevertheless, a single spectral clustering algorithm is affected by initialization, which results in unstable clustering results, especially when the data type and structure are more complicated. This paper proposes a novel approach of selective ensemble method based on spectral clustering to solve this issue. First of all, We obtain multiple spectral clustering results by random initialization parameters and generate candidate base clustering members. The normalized mutual information (NMI) is used to select high quality base clusterings from these candidate members as the input of clustering ensemble. Secondly, a new method of distance decision is devised as a consistency to evaluate the base clusterings. Finally, we introduce the cumulative matrix obtained by distance decision into the density peaks clustering method to gain the results of clustering ensemble. We implement the proposed method in UCI machine learning database that is commonly used in the field of clustering analysis, and experimental results demonstrate that our method can effectively improve the accuracy of single clustering algorithm.