An unsupervised feature-selection method (DSCOFS) that embeds two simultaneous sparsity constraints, an l2,0-norm and an l0-norm, into the principal component analysis framework, so the l2,0 term drops irrelevant and redundant features while the l0 term filters irregular noisy ones. The resulting nonconvex, nonsmooth problem is solved by a proximal alternating minimization scheme proven to converge globally to a stationary point, improving clustering accuracy and mutual information over single-sparsity methods.
This paper introduces DSCOFS, an unsupervised feature-selection method that adds a double-sparsity constraint to the principal component analysis framework: an l2,0-norm removes irrelevant and redund…