Robust semi-supervised concept factorization

Abstract

A robust semi-supervised concept factorization (RSSCF) method is proposed in this paper, which not only makes good use of the available label information, but also addresses noise and extracts meaningful information simultaneously. In the proposed method, a constraint matrix is embedded into the basic concept factorization model to guarantee data with the same label share the same new representation. We utilize L _2,1 -norm on both loss function and regularization, thus this new model is not sensitive to outliers and the L _2,1 -norm regularization helps select useful information with joint sparsity. An efficient and elegant iterative updating scheme is also introduced with convergence and correctness analysis. Simulations are given to illustrate the effectiveness of our proposed method.