Feature Maximization Based Clustering Quality Evaluation: A Promising Approach

Abstract

Feature maximization is an alternative measure, as compared to usual distributional measures relying on entropy or on Chi-square metric or vector-based measures, like Euclidean distance or correlation distance. One of the key advantages of this measure is that it is operational in an incremental mode both on clustering and on traditional classification. In the classification framework, it does not presents the limitations of the aforementioned measures in the case of the processing of highly unbalanced, heterogeneous and highly multidimensional data. We present a new application of this measure in the clustering context for setting up new cluster quality indexes whose efficiency ranges for low to high dimensional data and that are tolerant to noise. We compare the behaviour of these new indexes with usual cluster quality indexes based on Euclidean distance on different kinds of test datasets for which ground truth is available. Proposed comparison clearly highlights the superior accuracy and stability of the new method.