Measures for unsupervised fuzzy-rough feature selection

For supervised learning, feature selection algorithms attemptto maximise a given function of predictive accuracy.This function usually considers the ability of feature vectorsto reflect decision class labels. It is therefore intuitive to retainonly those features that are related to or lead to thesedecision classes. However, in unsupervised learning, decisionclass labels are not provided, which poses questionssuch as; which features should be retained? and, why notuse all of the information? The problem is that not all featuresare important. Some of the features may be redundant,and others may be irrelevant and noisy. In this paper, somenew fuzzy-rough set-based approaches to unsupervised featureselection are proposed. These approaches require nothresholding or domain information, can operate on realvalueddata, and result in a significant reduction in dimensionalitywhilst retaining the semantics of the data.