Gene clustering with partition around mediods algorithm based on weighted and normalized mahalanobis distance

Abstract

The partition around medoids (PAM) algorithm is a robust and flexible unsupervised learning algorithm that depends on the underlying distance and default distance metric is Euclidean distance. The PAM algorithm is more efficient than K-Mean since medoids assess a minimum distance from the other objects. In this study, we have integrated the Mahalanobis distance with PAM algorithm, since Mahalanobis distance has been defined in cluster analysis for different applications and is used to overcome the problem of scaling and correlation with Euclidean distance. However, the performance of PAM algorithm based on Mahalanobis distance was found to be inadequate when employed on selected microarray expression datasets, which were pre-processed prior to the analyses. We proposed an enhanced PAM algorithm based on the weighted and normalized Mahalanobis distance, and the results obtained using our proposed algorithm reveal an ultimate cluster solution for selected microarray datasets. The algorithms were evaluated using the Dunn's validity index.