Dimensionality Reduction with Unsupervised Ensemble Learning Using K-Means Variants

Feature selection aims to diminish dimensionality for construct comprehensible learning models with good generalization performance. Feature selection methods are mostly studied independently according to the type of learning: supervised or unsupervised. This paper describes a novel feature selection algorithm for unsupervised clustering, that combines two clustering ensembles methods such as bagging and random subspace method using K-means variants to unlabeled data that estimates the out-of-bag feature importance from an ensemble of partitions. Every partition is constructed using a various bootstrap samples and a random subset of the features. The principal idea of the proposed unsupervised feature selection method is to search for a subset of all features such that the clustering algorithm trained on this feature subset can reach the most identical clustering solution to the one acquired by a set learning method. Experiments are performed on different known data sets for validating our proposed method. The results are promising and competitive with various representative algorithms.