Group-weighted Subspace Clustering for Multivariate Compositional Data

Compositional data which contain relative information occur frequently in many practical scenarios.Promising achievements have been made in clustering for single compositional data variable, yet few works are available for multivariate compositional data clustering, let alone the importance measurement of compositional data variables.In this paper, we develop a novel clustering method for multivariate compositional data which integrates the self-expressive property and the group-weighted strategy.The selfexpressive property is generalized to multivariate compositional data analysis, upon which the coefficient matrix is obtained to reflect the global structure of data.Moreover, considering the importance of each compositional data variable is different during the process of clustering, we propose a group-weighted strategy to adaptively learn appropriate weights for them.Experimental results on a practical dataset demonstrate the effectiveness of the proposed method.