Clustering Ensemble for Unsupervised Feature Selection

A new feature selection algorithm for unsupervised learning is proposed. It is based on the assumption that, in absence of class labels, the clustering ensemble result can be employed as a heuristic to guide the feature selection. Therefore, a modified RReliefF algorithm is then used to assign the rankings for every feature. The main advantage of the proposed unsupervised feature selection algorithm in comparison to conventional schemes is that it is dimensionality unbiased. Our experiments with several data sets demonstrate that the proposed algorithm is able to detect completely irrelevant features and to remove some additional features without significantly hurting the performance of the clustering algorithm.