Robust discriminative non-negative matrix factorization

摘要: Traditional non-negative matrix factorization (NMF) is an unsupervised method that represents data by a part-based dictionary and codes. Recently, the NMF has been extended to discriminative ones for classification problems. However, these methods may become inefficient when outliers are presented in data, e.g. mislabeled samples, because usually deviate from normal samples one class would perturb dictionary. In this paper, we propose novel method, called robust (RDNMF), reduce effect of improve strength. The RDNMF learns each class, contains two parts: part outlier part. parts obtained minimizing cosine similarity between classes. codes on required be sparse so most can modeled part, without large influence over part.The final concatenating all classes, sample, as well test coding with Experimental comparisons existing learning MNIST, PIE, Yale B ORL demonstrate effectiveness robustness our approach.