Feature selection in clustering with constraints: Application to active exploration of music collections

Constrained clustering has been developed to improve clustering methods through pairwise constraints. Although the constraints are enhancing the similarity relations between the items, the clustering is conducted in the static feature space. In this paper we embed the information about the constraints to a feature selection procedure, that adapts the feature space regarding the constraints. We propose two methods for the constrained feature selection: similarity-based and constrained-based. We apply the constrained clustering with embedded feature selection for the active exploration of music collections. Our experiments show that proposed feature selection methods improve the results of the constrained clustering.