nonnegative matrix factorization via rank-one downdate
DESCRIPTION
Nonnegative Matrix Factorization via Rank-one Downdate. Nonnegative Matrix Factorization. -2.19. -3.19. -0.02. 1.02. 2 by 1965. 560 by 1965. 560 by 2. 20 by 28. 2 by 1. 2 by 1. 20 by 28. Singular Value Decomposition (SVD). History. History. History. History. History (Algorithms). - PowerPoint PPT PresentationTRANSCRIPT
Nonnegative Matrix Factorization via Nonnegative Matrix Factorization via Rank-one DowndateRank-one Downdate
Nonnegative Matrix Nonnegative Matrix FactorizationFactorization
560 by 1965
560 by 2
2 by 1965
20 by 28 20 by 28
-2.19
-0.02
-3.19
1.02
2 by 12 by 1
Singular Value Decomposition Singular Value Decomposition (SVD)(SVD)
HistoryHistory
HistoryHistory
HistoryHistory
HistoryHistory
History (Algorithms)History (Algorithms)
History (Algorithms)History (Algorithms)
First observationFirst observation
Power methodPower method
• Computes the leading singular vectors/value (or eigenvector/value) of a matrix
:)d(powermetho,, Avu
1
2 while not converged
uA
uA
uAv
AvAv
u
T
T
T
ones of vectorv
3
4
5
6 end
ExampleExample
• Computes the leading singular vectors/value (or eigenvector/value) of a matrix
:)d(powermetho,, Avu
2 while not converged
uA
uA
uAv
AvAv
u
T
T
T
ones of vectorv
3
4
5
6 end
1
2
3
4
5 for all set
6 end for
Naive approach to NMF using this Naive approach to NMF using this observationobservation
Without step 5, this will simply compute the SVD (Jordan's algorithm, Camille Jordan 1874. )
:)nmf(, AHW
ki : for 1)d(powermetho],,[ Avu
TTii vHuW ,
TvuAA 0jiA ,0jiA ,
Second observationSecond observation
Modified power iterationModified power iteration
:)d(powermethomodified,, Avu
1
2 while not converged
uA
uA
uAv
AvAv
u
T
T
T
method power using initial find v
3
4
5
6
7
8 end
mean smaller withcluster the remove and cluster u
mean smaller withcluster the remove and cluster v
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
A =
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
0.14 0.07 0.64 0.41 0.55v:
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
0.0 0.0 0.64 0.41 0.55v:
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
v: 0.0 0.0 0.64 0.41 0.55
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
u:
0.160.210.220.440.740.20
v: 0.0 0.0 0.64 0.41 0.55
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
u:
v: 0.0 0.0 0.64 0.41 0.55
0.00.00.0
0.440.740.20
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
u:
v: 0.0 0.0 0.64 0.41 0.55
0.00.00.0
0.440.740.20
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
u:
v: 0.0 0.0 0.60 0.28 0.59
0.00.00.0
0.440.740.20
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
u:
v:
0.00.00.0
0.440.740.20
0.0 0.0 0.60 0.28 0.59
Modified power iteration: DemoModified power iteration: Demo
Rank-1submatrix
Rank-1submatrix
u:
v:
0.00.00.0
0.440.740.20
0.0 0.0 0.60 0.28 0.59
Zero-out!Zero-out!
Modified power iteration: DemoModified power iteration: Demo
Anew =
Rank-1submatrix
Rank-one Downdata (R1D)Rank-one Downdata (R1D)
1
2
3
4
5 end for
:)nmf(, AHW
ki : for 1
)wermethod(modifiedpo],,[ Avu TT
ii vHuW ,
000 yxyx vuA and if Let ,
Objective functionObjective function
ApproxRankOneSubmatrix(A)ApproxRankOneSubmatrix(A)
Rank-one Downdata (R1D)Rank-one Downdata (R1D)
Rank-one Downdata (R1D)Rank-one Downdata (R1D)
A simple model for textA simple model for text
Generating a corpus in the Generating a corpus in the modelmodel
Theorem about textTheorem about text
LSILSI
R1DR1D
Experimental resultsExperimental results
• Information retrieval task– TDT corpus (pilot study, v. 1.3, 1997): news
articles
Identified topics (first few columns of ):W
Topic: president, clinton, house, white
Topic: bosnian, serbs, bosnia, serb, nato, sarajevo, air, bihac
Topic: haiti, military, aristide, haitian, troops, port, invasion, …
Topic: simpson, defense, judge, case, jury, trial, angeles, los,
court, …
Topic: bill, today, senate, republicans, house, congress,
republican, …
Experimental resultsExperimental results• For topic: “OJ Simpson trial” (simpson, defense, judge,
case, jury, trial)
Document 1Allen
Simpson
Judge
Defence
Marcus
Court
Witness
Testify
Nicole
gloves
Document 2Simpson
Judge
Defense
Statements
Opening
Jury
Legal
Yesterday
Prosecution
Marc
los
Document 3Simpson
Gloves
Case
Prosecution
Defense
Roger
Put
Fit
Today
Problems
jury
Document 4Simpson
Kaelin
Police
Defense
Chicago
Testimony
Night
Knapsack
Cnn
Blood
Witness
Theorem about imagesTheorem about images
Experimental resultsExperimental results
LSILSI
NMF-DIVNMF-DIV
R1DR1D
LSI
NMF_DIV
NMF_SC
R1D