A robust dimensionality reduction and matrix factorization framework for data clustering

摘要: Abstract Most existing Non-negative Matrix Factorization (NMF) related data clustering techniques directly decompose the original feature space while have not well considered fact that low dimensional is always embedded in high and can better reveal spatial distribution of data. In this letter, we propose a new matrix factorization model, which unites objectives dimensionality reduction simultaneously. proposed framework, based on actually executed subspace may provide more accurate reasonable solutions. Furthermore, use l2,1-norm instead conventional l2-norm to enhance results make framework robust noises outliers. Meanwhile, order preserve as much possible local similarity data, also employed an affinity with special learning introduce manifold into cluster indicator matrix. An optimization procedure Augmented Lagrangian Method (ALM) devised effectively solve problem explicitly show results. Experimental benchmark datasets different proprieties exhibit superior performance method.