Robust Nonnegative Matrix Factorization Yining Zhang 04-27-2012 1 Outline Review of nonnegative matrix factorization Robust nonnegative matrix factorization using L21-norm Robust nonnegative matrix factorization through sparse learning Further works 2 Outline Review of nonnegative matrix factorization Robust nonnegative matrix factorization using L21-norm Robust nonnegative matrix factorization through sparse learning Further works 3 Review of nonnegative matrix factorization 4 5 6 Clustering Interpretation 7 Outline Review of nonnegative matrix factorization Robust nonnegative matrix factorization using L21-norm Robust nonnegative matrix factorization through sparse learning Further works 8 Robust nonnegative matrix factorization using L21-norm 9 Shortcoming of Standard NMF 10 L21-norm 11 From Laplacian noise to L21 NMF 12 Illustration of robust NMF on toy data 13 14 Illustration of robust NMF on real data 15 16 Computation algorithm for L21NMF 17 Convergence of the algorithm Theorem 1. (A) Updating G using the rule of Eq.(17) while fixing F, the objective function monotonically decreases. (B) Updating F using the rule of Eq.(16) while fixing G, the objective function monotonically decreases. 18 Updating G 19 Correctness of the algorithm Theorem 7. At convergence, the converged solution rule of Eq.(17) satisfies the KKT condition of the optimization theory. 20 A general trick about the NMF KKT condition Updating formula Auxiliary function Prove monotonicity 21 22 Experiments on clustering 23 24 Outline Review of nonnegative matrix factorization Robust nonnegative matrix factorization using L21-norm Robust nonnegative matrix factorization through sparse learning Further works 25 Robust nonnegative matrix factorization through sparse learning 26 Motivation Motivated by robust pca 27 Optimization 28 Experimental results-1 A case study 29 Experimental results 2Face clustering on Yale 30 Experimental results 3Face recognition on AR 31 Outline Review of sparse learning Efficient and robust feature selection via joint l2,1-norm minimzation Exploiting the entire feature space with sparsity for automatic image annotation Further works 32 Future works-1 (1) Direct robust matrix factorization for anomaly detection. 2011 ICDM. 33 Future works-2 34 Reference [1]Deguang Kong, Chris Ding, Heng Huang. Robust nonnegative matrix factorization using L21-norm. CIKM 2011. [2]Lijun Zhang, Zhengguang Chen, Miao Zheng, Xiaofei He. Robust nonnegative matrix factorization. Front. Electr. Eng.China 2011. [3]Chris Ding, Tao Li, Michael I.Jordan. Convex and Semi-nonnegative matrix factorization. IEEE T.PAMI, 2010.. 35 Thanks! Q&A 36