MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
MACHINE LEARNING TECHNIQUES - LASA
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
27<br />
The CCA algorithm consists thus in finding the optimum of ρ under the above two constraints.<br />
This again can be resolved through Lagrange and gives:<br />
C C C w<br />
−1 2<br />
xy yy yx x xx x<br />
= λ C w<br />
(2.18)<br />
with λ the Lagrange multiplier of the constraint.<br />
This is a generalized eigenproblem of the form Ax = λBx<br />
. If C is invertible (which is the case if<br />
xx<br />
all the columns of X are independent), then the above problem reduces to a single symmetric<br />
−<br />
eigenvector problem of the form B 1 Ax= λx.<br />
2.2.1 CCA for more than two variables<br />
Given sets of K multivariate random variables<br />
k<br />
k<br />
X with dimension , 1,...,<br />
M× n k = K(note that<br />
k<br />
while each variate can have different dimensions n , k = 1,..., K , they must all have the same<br />
number of observed instances M). The generalized CCA problem consists then in determining the<br />
= , with p= min k = 1.. K<br />
n k<br />
, such that:<br />
k k<br />
W w<br />
=<br />
set { i }<br />
i 1...<br />
p<br />
min<br />
W<br />
∑<br />
XW<br />
l≠ k, k, l=<br />
1,..., K F<br />
k<br />
( )<br />
k k l l<br />
T<br />
k<br />
s.t. W C W = I, ∀ k = 1,.., K,<br />
kk<br />
− XW<br />
k l<br />
wiCklw j<br />
= 0, ∀ i, j = 1,..., p, i ≠ j.<br />
(2.19)<br />
Where is the Frobenius norm (Euclidean distance for matrices). The above consists of K<br />
F<br />
minimization under constraint problems. Unfolding the objective function, we have the sum of the<br />
squared Euclidean distances between all of the pairs of the column vectors of the<br />
k k<br />
matrices XW, k= 1,..., K . This problem can be solved by using singular value decomposition<br />
for arbitrary K.<br />
2.2.2 Limitations<br />
CCA is dependent on the coordinate system in which the variables are described, so even if there<br />
is a very strong linear relationship between two sets of multidimensional variables, depending on<br />
the coordinate system used, this relationship might not be visible as a correlation. Kernel CCA<br />
has been proposed as an extension to CCA, where data are first projected into a higher<br />
dimensional representation. KCCA is covered in Section 5.4 of these Lecture Notes.<br />
© A.G.Billard 2004 – Last Update March 2011