ÁREA DE EVALUACIÓN 4: GOBERNABILIDAD DEMOCRÁTICA
5.3.4. Área 4: Capacidades de gestión Gobernabilidad Democrática
Similar to PLS, the two-block CCA finds linear combinations of rows ofX
1andX
2maximizing
sequentially from the optimization problem:
{~a
(i)1, ~a
(i)2}=
arg max
~a1∈Rd1,~a2∈Rd2
h~a
|1X
1, ~a
|2X
2i
subject to the constraints:
k~a
|1X
1k= 1,k~a
|2X
2k= 1
h~a
|1X
1,(~a
(j) 1)
|X
1i= 0,h~a
|2X
2,(~a
(j) 2)
|X
2i= 0, j
= 1,· · ·
, i−1.
(4.10)
This form makes the relationship between (4.5) and (4.10) clear (differing only constraining the
loading loadingsk~a
kk= 1 versus the scoresk~a
|kX
kk= 1) and is equivalent to the usual formulation
of optimizing the correlation.
There is an important relationship between CCA and PAA (Hotelling, 1936; Bj¨orck and Golub,
1973), i.e., if
ρ
i=h(~a
(i)1)
|X
1,(~a
2(i))
|X
2i
is the
ith canonical correlation,
ρ
i= cos(θ
i), where
θ
iis
theith principal angle between row spaces ofX
1and
X
2. The principal vector pairs{~x
1,i, ~x
2,i}=
{X
|1~a
(i)1,X
2|~a
(i)2}are often obtained through SVD ofV
X|1
V
X2. In particular, let~u
X1,i, ~u
X2,ibe the
ith left and right singular vectors ofV
|X1
V
X2. Then, theith pair of principal vectors are
~
x
1,i=V
X1~u
X1,i,
~x
2,i=V
X2~u
X2,i.
An issue with CCA of high-dimensional data is related to the fact that CCA is interested in
the canonical vectors~a
irather than the principal vectors
~x
i. In particular, when
d
1> n, d
2> n,
the values of~a
iin (4.10) are not identifiable due to the singularity of
X
1X
|1and
X
2X
|2. Several
approaches have been taken to solve this problem. One approach is to use the Moore-Penrose pseudo
inverse in place of the inverse of
X
1X
|1and
X
2X
|2. A second approach is to add a ridge penalty
on
X
1X
|1and
X
2X
|2(Vinod, 1976). A third approach called
penalized CCA
is to add penalty
functions on
{~a
(i)1, ~a
(i)2}, such as an
`
1penalty (Parkhomenko et al., 2007; Lˆe Cao et al., 2009),
an elastic net (Waaijenborg et al., 2008) or a fused lasso (Witten et al., 2009). Another approach
called
diagonal penalized CCA
is to replace
X
1X
|1and
X
2X
2|by diag(X
1X
|1) and diag(X
2X
|2)
(Parkhomenko et al., 2009; Witten et al., 2009).
Another important issue with CCA, which is directly related to AJIVE, is that when
d
1>
n, d
2> n, CCA is generally driven by noise. Lee (2007); Samarov (2009); Lee (2016) study the
inconsistency phenomenon in this case. One solution to this issue, used by AJIVE and COBE, is to
replace
X
kby its low rank approximation ˜A
k,k= 1,2. Theith principal vectors are~p
i= ˜V
1~u
1,i,
~
q
i= ˜V
2~u
2,i, where~u
j,iis theith singular vector of ˜U
iof the SVD of ˜V
|1V˜
2respectively.
As described in Section 4.3, AJIVE uses an equivalent principal angle calculation based on
SVD of
M= [ ˜V
1,V˜
2]
|=U
MΣ
MV
M|(Miao and Ben-Israel, 1992). AJIVE uses the transpose of
the
ith right singular vector,
~v
M,i, as the estimated
ith basis vector of the joint space, provided
that the
ith principal angle is smaller than the threshold derived in Section 3.3.3.1. Moreover, if
the
ith principal angle has a value distinct from other principal angles, then the
ith left singular
vector of
Mcan be written as
~u
M,i= [~u
|1,i, ~u
| 2,i
]
|/
√
2. Consequently
~
v
M,i=
1
σ
M,iM
|~u
M,i=
1
√
2σ
M,i( ˜V
1~u
1,i+ ˜V
2~u
2,i) =
1
√
2σ
M,i(~p
i+~q
i).
This shows that the AJIVE direction~v
M,iis the scaled sum (essentially the average) of theith pair
of principal vectors.
CCA applied to the low rank approximations ˜A
kand AJIVE are therefore closely related.
However, AJIVE provides one joint vector per two distinct principal vectors that by the virtue of
being an average should be a better estimate of the joint space than either of the principal vectors.
More importantly, AJIVE uses a theoretically sound threshold of the principal angles that allows
us to segment individual and joint variation.
multiset CCA
The AJIVE formulation allows for a natural extension to multi-block situations.
Several approaches of Multiset Canonical Correlation Analysis (mCCA) have been developed as
extensions of CCA (Horst, 1961; Kettenring, 1971; Nielsen, 2002). There is no general consensus
on which of these extensions is preferable. We point out that AJIVE is closely related to one of
the mCCA approaches discussed in (Nielsen, 2002).
This version of mCCA is defined using the optimization problem for the
ith set of canon-
ical vectors
{~a
(i)1,· · ·, ~a
(i)K}
and corresponding principal vectors (also called canonical variables)
{X
|1~a
(i)1,· · ·,X
|K~a
(i)K}:
{~a
(i)1,· · ·
, ~a
(i)K}= arg max
~a1,···,~aK
X
1≤k,l≤K
h~a
|kX
k, ~a
|lX
li
subject to the constraints:
K
X
k=1k~a
|kX
kk
22= 1,
h~a
|kX
k,(~a
(j)k)
|X
ki= 0,
k= 1,· · ·, K,
j
= 1,· · ·
, i−1.
(4.11)
Notice that the constraint in (4.11) is different than the perhaps more naturalk~a
|kX
kk
22= 1 for all
k.
If the
ith singular value corresponding to the AJIVE direction~v
M,ihas a value distinct from
other singular values in the AJIVE SVD, then calculations similar to the two block case show that
theith basis vector of the joint space from AJIVE
~v
M,i=
1
σ
M,i KX
k=1X
|k~a
(i)kis again a scaled sum (essentially average) of the corresponding canonical variables.
In document
La gestión municipal, propuesta de mejora para la Municipalidad Provincial de Bolívar, 2017
(página 99-104)