Improving Clustering and Cluster Validation with Missing Data Using Distance Estimation Methods

Missing data introduces a challenge in the field of unsupervised learning. In clustering, when the form and the number of clusters are to be determined, one needs to deal with the missing values both in the clustering process and in the cluster validation. In the previous research, the clustering algorithm has been treated using robust clustering methods and available data strategy, and the cluster validation indices have been computed with the partial distance approximation. However, lately special methods for distance estimation with missing values have been proposed and this work is the first one where these methods are systematically applied and tested in clustering and cluster validation. More precisely, we propose, implement, and analyze the use of distance estimation methods to improve the discrimination power of clustering and cluster validation indices. A novel, robust prototype-based clustering process in two stages is suggested. Our results and conclusions confirm the usefulness of the distance estimation methods in clustering but, surprisingly, not in cluster validation.