Discovery of Linear Acyclic Models Using Independent Component Analysis
•
3 likes•5,712 views
This document discusses the discovery of linear acyclic models from non-experimental data using independent component analysis (ICA). It describes how existing methods assume Gaussian disturbances, producing equivalent models, whereas the proposed LiNGAM approach assumes non-Gaussian disturbances. This allows identifying the connection strengths and structure without equivalent models. The LiNGAM algorithm estimates the matrix B using ICA and post-processing, finds a causal order, and prunes non-significant edges. Examples show LiNGAM can correctly estimate networks and the document concludes it is an important topic with code available online.
1 of 16
Download to read offline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
More Related Content
Discovery of Linear Acyclic Models Using Independent Component Analysis
1. Created by S.S. in Jan 2008
Discovery of Linear Acyclic Models
Using Independent Component Analysis
Shohei Shimizu, Patrik Hoyer,
Aapo Hyvarinen and Antti Kerminen
LiNGAM homepage: http://www.cs.helsinki.fi/group/neuroinf/lingam/
2. 2
Independent Component Analysis
(Hyvarinen et al., 2001)
x As
m 㬍1 m 㬍n n 㬍1
• A is unknown, cov(s)=I
– Typically, A is square (m=n)
i s
• are assumed to be non-Gaussian and mutually
independent
• Non-Gaussian and independence are
the key assumptions in ICA
• Estimable including the rotation (Comon, 1994)
3. 3
Linear acyclic models
(Bollen, 1989; Pearl, 2000; Spirtes et al., 2000)
• Continuous variables
• Directed acyclic graph (DAG)
• The value assigned to each variable is a linear
combination of those previously assigned, plus a
disturbance term ei
• Disturbances (errors) are independent and have non-zero
variances
x Bx e i
i ij j e x b x
k j k i
( ) ( )
or
4. 4
Our goal
• We know
– Data X is generated by
• We do NOT know
x Bx e
– Connection strengths: B
– Order: k(i)
– Disturbances: ei
• What we observe is data X only
• Goal
– Estimate B and k using data X only!
5. 5 Can we recover the original
network?
Original network
?
6. 6
Can we recover the original
network using ICA?
Yes!
Original network Estimated network
7. 7
Discovery of linear acyclic models
from non-experimental data
• Existing methods (Bollen,1989; Pearl, 2000; Spirtes et al., 2000)
– Gaussian assumption on disturbances ei
– Produce many equivalent models
• Our LiNGAM approach (Shimizu et al, UAI2005, 2006 JMLR)
– Replace Gaussian assumption by non-Gaussian assumption
– Can identify the connection strengths and structure
Gaussnianity
Non-Gaussianity
Equivalent models
No Equivalent models
x1 x2
x3
x1 x2
x3
x1 x2
x3
x1 x2
x3
x1 x2
x3
x1 x2
x3
8. 8 Linear Non-Gaussian Acyclic Models
(LiNGAM)
• As usual, linear acyclic models,
but disturbances are assumed to be
non-Gaussian:
i ij j e x b x
• Examples
x Bx e i
k j k i
( ) ( )
or
i e
9. Outline of LiNGAM algorithm 9
x2
x3
x4
x1
Non-significant edges
x2
x3
x4
x1
1. Estimate B by ICA
+ post-processing
2. Find an order k(i) (DAG)
3. Prune non-significant edges
x2
x3
x4
x1
10. 10 Key ideas
i x
• Observed variables are linear combinations of
non-Gaussian independent disturbances
x Bx
e
( )1
x I B e
Ae
– The classic case of ICA (Independent Component Analysis)
• Permutation indeterminacy in ICA can be solved
– Can be shown that the correct permutation is the only one which
has no zeros in the diagonal (Shimizu et al., 2005; 2006)
• Pruning edges can be done by many existing methods
– Imposing sparseness, testing, model fit etc.
i e
11. 11
Examples of estimated networks
• All the edges correctly identified
• All the connection strengths approximately correct
Original network Estimated network
12. 12 What kind of mistakes LiNGAM might make?
• One falsely added edge (x1x7, -0.019)
• One missing edge (x1x6, 0.0088)
Original network Estimated network
13. 13
Real data example 1
• Galton’s height data (Galton, 1886)
– x1: child height
– x2: `midparent’ height
• (father’s height + 1.08 mother’s height)/2
– 928 observations
• Estimates:
x
1
0 0.67
• Estimated direction:
15. e
1
2
x
1
2
2
0.012 0
e
x
x
1 2 x x child midparent
16. 14
Real data example 2
• Fuller’s corn data (Fuller, 1987)
– x1: Yield of corn
– x2: Soil Nitrogen
– 11 observations
• Estimates:
x
1
0 0.34
• Estimated direction:
19. e
1
2
x
1
2
2
0.014 0
e
x
x
1 2 x x Yield Nitrogen
20. 15
Conclusions
• Discovery of linear acyclic models from non-experimental data is an
important topic of current research
• A common assumption is linear-Gaussianity, but this leads to a
number of indistinguishable models
• A non-Gaussian assumption allows all the connection strengths and
structure of linear acyclic models to be identified
• Basic method: ICA + permutation + pruning
• Matlab/Octave code: http://www.cs.helsinki.fi/group/neuroinf/lingam/
21. 16
References
• S. Shimizu, A. Hyvärinen, Y. Kano, and P. O. Hoyer
Discovery of non-gaussian linear causal models
using ICA. In Proceedings of the 21st Conference on
Uncertainty in Artificial Intelligence (UAI2005), pp. 526-
533, 2005.
• S. Shimizu, P. O. Hoyer, A. Hyvärinen, and A. Kerminen.
A linear non-gaussian acyclic model for causal
discovery. Journal of Machine Learning Research, 7:
2003--2030, 2006.