Empricial Study of Cluster Evaluation Metrics

A wide range of abstract characteristics of partitions have been proposed for cluster evaluation. We empirically evaluated the performance of these metrics for flow cytometry data and found that the set-matching metrics perform closest to human. Clustering is an increasingly popular module in data processing applications. Many clustering algorithms have been developed and many more are anticipated to emerge in the future. Thus, methods for assessing the performance of a clustering algorithms are in great demand. Such methods assess the performance of a clustering algorithm by computing a quality score of the solution against a ground truth partition, usually designed by a human expert. A wide range of these criterion have been proposed [9]. Evaluating clustering algorithms heavily relies on the chosen quality score however, it is not practical to study the performance of these metrics in a domain-independent way [3]. In this paper we aim to empirically evaluate the available metrics to find the best metric for comparing clustering solutions against ground truth partitions for flow cytometry (FCM) applications. This work was motivated in part by the challenges we faced in choosing the best clustering comparison metric for the FlowCap project. FlowCap is an international open project designed to provide an objective way to compare and evaluate FCM data clustering methods, and also to establish guidance about appropriate use and application of these methods (for more information visit http://flowcap.flowsite.org/).