A Hierarchical Clustering Algorithm Based on Silhouette Index for Cancer Subtype Discovery from Omics Data

Abstract

Cancer subtype discovery from omics data requires techniques to estimate the number of natural clusters in the data. Automatically estimating the number of clusters has been a challenging problem in Machine Learning. Using clustering algorithms together with internal cluster validity indexes have been a popular method of estimating the number of clusters in biomolecular data. We propose a Hierarchical Agglomerative Clustering algorithm, named SilHAC, which can automatically estimate the number of natural clusters and can find the associated clustering solution. SilHAC is parameterless. We also present two hybrids of SilHAC with Spectral Clustering and K-Means respectively as components. SilHAC and the hybrids could find reasonable estimates for the number of clusters and the associated clustering solution when applied to a collection of cancer gene expression datasets. The proposed methods are better alternatives to the 'clustering algorithm -internal cluster validity index' pipelines for estimating the number of natural clusters.

Development of a global batch clustering with gradient descent and initial parameters in colour image classification

“…In the first step, Algorithm 1 is used to generate the initial parameters. The effectiveness of the initialisation is then evaluated in comparison of another three typical existing estimation methods, namely Ward's method [8, 51], the silhouette index [27] and Sujatha's method [20]. The parameters by four estimation methods are refined with the original batch clustering using K ‐means.…”

“…The initial parameters are critical for the clustering process as justified. To evaluate theeffectiveness of Algorithm 1, another three existing methods to create initialparameters for K ‐means algorithm are selected: Ward's method [8, 51],the silhouette index [27] and Sujatha'smethod [20]. The Ward's method is anagglomerative hierarchical procedure with the object function of the error sum ofsquare measure.…”

“…Hence, the spatial location of the cluster in the coordinate space has beenneglected. Even the silhouette index has been proposed to merge the clusters basedon the other method such as the Ward framework [27]. The merging of the clustering with high‐silhouette index is stillbased on the formed local optima.…”