Subspace clustering of high-dimensional data: a predictive approach