A spectral clustering algorithm based on attribute fluctuation and density peaks clustering algorithm

Spectral clustering (SC) has become a popular choice for data clustering by converting a dataset to a graph structure and then by identifying optimal subgraphs by graph partitioning to complete the clustering. However, k-means is taken at the clustering stage to randomly select the initial cluster centers, which leads to unstable performance. Notably, k-means needs to specify the number of clusters (prior knowledge). Second, SC calculates the similarity matrix using the linear Euclidean distance, losing part of the effective information. Third, real datasets usually contain redundant features, but traditional SC does not adequately address multi-attribute data. To solve these issues, we propose an SC algorithm based on the attribute fluctuation and density peaks clustering algorithm (AFDSC) to improve the clustering accuracy and effect. Furthermore, to verify the idea of the AFDSC algorithm, we extract the attribute fluctuation factor and propose a histogram clustering algorithm based on attribute fluctuation (AFHC) divorced from spectral clustering. Experimental results show that both the AFDSC algorithm and AFHC algorithm have achieved better performance on fifteen UCI datasets compared with other clustering algorithms.