Robust Dual Concept Factorization for Data Clustering

Concept factorization approximates the original data matrix with concepts and coefficients by minimizing the squared loss and has been widely used for data representation and clustering. The graph regularization has also been adopted on both sides of the factorization. In this paper, we propose the novel robust dual concept factorization method to improve the robustness against outliers and noises. In particular, we take the dual concept factorization model and use the corr-entropy induced metric to guide the data reconstruction and the corresponding dual graph regularization. As a result, the affect of the outliers can be alleviated from both data reconstruction and dual regularization. Performance also validates the method proposed to group data on reference data sets.