Nonnegative Matrix Factorization via Rank-one Downdate Ali Ghodsi Department of Statistics and Actuarial Science David R. Cheriton School of Computer Science University of Waterloo Joint work with Stephen Vavasis and Michael Biggs
Nonnegative Matrix Factorization
2 by 1965 560 by 1965 560 by 2 -2.19 -0.02 -3.19 1.02 20 by 28 20 by 28 2 by 1 2 by 1
Singular Value Decomposition (SVD)
History
History
History
History
History (Algorithms)
History (Algorithms)
First observation
Power method Computes the leading singular vectors/value (or eigenvector/value) of a matrix 1 2 while not converged 3 4 5 6 end
Naive approach to NMF using this observation 1 2 3 4 5 for all set 6 end for Without step 5, this will simply compute the SVD (Jordan's algorithm, Camille Jordan 1874. )
Rank-one Downdata (R1D)
Objective function
ApproxRankOneSubmatrix(A)
Modified power iteration: Demo Rank-1 submatrix A = Rank-1 submatrix
Modified power iteration: Demo v: 0.14 0.07 0.64 0.41 0.55 Rank-1 submatrix Rank-1 submatrix
Modified power iteration: Demo v: 0.0 0.0 0.64 0.41 0.55 Rank-1 submatrix Rank-1 submatrix
Modified power iteration: Demo v: 0.0 0.0 0.64 0.41 0.55 Rank-1 submatrix Rank-1 submatrix
Modified power iteration: Demo v: 0.0 0.0 0.64 0.41 0.55 0.16 0.21 0.22 0.44 0.74 0.20 Rank-1 submatrix u: Rank-1 submatrix
Modified power iteration: Demo v: 0.0 0.0 0.64 0.41 0.55 0.0 0.44 0.74 0.20 Rank-1 submatrix u: Rank-1 submatrix
Modified power iteration: Demo v: 0.0 0.0 0.64 0.41 0.55 0.0 0.44 0.74 0.20 Rank-1 submatrix u: Rank-1 submatrix
Modified power iteration: Demo v: 0.0 0.0 0.60 0.28 0.59 0.0 0.44 0.74 0.20 Rank-1 submatrix u: Rank-1 submatrix
Modified power iteration: Demo v: 0.0 0.0 0.60 0.28 0.59 0.0 0.44 0.74 0.20 Rank-1 submatrix u: Rank-1 submatrix
Modified power iteration: Demo v: 0.0 0.0 0.60 0.28 0.59 0.0 0.44 0.74 0.20 Rank-1 submatrix u: Rank-1 submatrix Zero-out!
Modified power iteration: Demo Rank-1 submatrix Anew =
Rank-one Downdata (R1D)
A simple model for text
Generating a corpus in the model
Theorem about text
LSI
R1D
Theorem about images
Experimental results
LSI
NMF-DIV
R1D
LSI
NMF_DIV
R1D
Thank you!