Feature Selection based Semi-Supervised Subspace Clustering

Clustering is the process which is used to assign a set of n objects into clusters(groups). Dimensionality reduction techniques help in increasing the accuracy of clustering results by removing redundant and irrelevant dimensions. But, in most of the situations, objects can be related in different ways in different subsets of the dimensions. Dimensionality reduction tends to get rid of such relationship information and generate clusters which do not fully reflect the real cluster’s properties. Subspace clustering preserves such relationships by detecting all clusters in all subspaces. The accuracy of the subspace clustering results can be improved by making use of semi-supervised learning method. But finding subspaces by considering all input dimensions may decrease the clustering accuracy. This paper proposes a feature selection based semi-supervised subspace clustering method which applies feature selection in the beginning to eliminate unnecessary dimensions. Later, subspace clustering can be performed on the resulting dataset. This approach tends to improve the accuracy of resulting clusters since subspace clustering is performed on a reduced dataset. Experimental results show that the proposed method produces high quality clusters than semi-supervised subspace clustering algorithm. General Terms Dimensionality Reduction, Clustering.