Clustering quality based feature selection method
Artur Klepaczko, Andrzej Materka · Machine Graphics & Vision International Journal archive · 2004
This paper focuses on the problem of dimensionality reduction for objects described by a large number of features. The emphasis is put on the issues of grouping unlabelled data sets, where information about class-membership of observations is unavailable. Commonly used feature extraction methods for unsupervised classification tasks (such as PCA) are not applicable when information necessary for partitioning of the data set is not represented by the data structure as a whole, but is hidden in a limited number of features only. Thus, we propose a novel technique for choosing the best discriminative data features in an unsupervised manner. Our approach is based on data clustering and on clustering quality measures. The method is straightforward but proved perceptive and efficient. Since the research was primarily motivated by the specific problem of classifying MRI data, performance of the constructed algorithm is studied in application to textured image analysis.