3.7. MULTI-MODAL TRANSDUCTIVE LOW-RANK LEARNING 45
where O
Z
H D ˇw
t
.w
T
t
Z
t
M y
T
/M
T
C 2ˇw
t
w
T
t
Z
t
L
t
X
T
O
W
t
.
O
W
T
t
X
O
W
T
t
XZ
t
E
t
C
1
t
Y
1;t
/ is the partial derivative H.Z; E; W; Y
1
; Y
2
; / with respect to Z and
t
D k
O
W
T
t
Xk
2
F
.
e problem in Eq. (3.38) is a standard nuclear norm minimization problem, which can
be approximately solved by the singular value thresholding (SVT) algorithm [14]. Specifically,
suppose that the singular vector decomposition of Z
t
O
Z
H of rank r is
Z
t
O
Z
H D P†Q
T
; † D diag
f
ı
i
g
r
iD1
; (3.39)
where P and Q are left-singular and right-singular matrices with orthogonal columns and †
is a rectangular diagonal matrix with non-negative real numbers ı
i
on the diagonal. en, the
optimal solution Z is Z
tC1
D D
1=
t
t
.Z
t
O
Z
H /. For each 1=
t
t
0, the soft-thresholding
operator D
1=
t
t
.Z
t
O
Z
H / is defined as [14]:
8
<
:
D
1=
t
t
.Z
t
O
Z
H / D P†
1=
t
t
C
Q
T
;
†
1=
t
t
C
D diag.f.ı
i
1=
t
t
/
C
g/;
(3.40)
where t
C
is the positive part of t , namely, t
C
D max.0; t/.
For E: We can obtain the optimization of E with fixed w; Z; and W as follows:
E
tC1
D arg min
E
2
O
W
T
t
X
O
W
T
t
XZ
tC1
E
2
F
C
D
Y
1;t
;
O
W
T
t
X
O
W
T
t
XZ
tC1
E
E
C
k
E
k
1
D arg min
E
t
k
E
k
1
C
1
2
E
O
W
T
t
U
tC1
Y
1;t
=
t
2
F
;
(3.41)
where U
tC1
D X XZ
tC1
is defined for simplicity. e optimization of Eq. (3.41) can be solved
by using the shrinkage operator [90].
For w: We can optimize w with fixed E; Z; and W as follows:
w
tC1
D arg min
w
1
2
y .Z
tC1
M/
T
w
2
2
C w
T
ZLZ
T
w C
˛
ˇ
k
w
k
2
2
: (3.42)
e above problem is actually the well-known ridge regression, whose optimal solution is
w
tC1
D
Z
tC1
MM
T
Z
T
tC1
C 2Z
tC1
LZ
T
tC1
C
2˛
ˇ
I
1
Z
tC1
My.
For W: By setting the derivative of L regarding W to zero, we have
O
S
O
W
tC1
.Y
2;t
C Y
T
2;t
/ 2ıˇS
O
W
tC1
C
t
U
tC1
U
T
tC1
O
W
tC1
D U
tC1
E
T
tC1
U
tC1
Y
T
1;t
:
(3.43)
en, W
tC1
can be optimized by solving the Lyapunov equation.
Moreover, the Lagrange multipliers Y
1
and Y
2
are updated by the following scheme:
8
<
:
Y
1;tC1
D Y
1;t
C
t
O
W
T
tC1
U
tC1
E
tC1
;
Y
2;tC1
D Y
2;t
C
t
O
W
T
tC1
O
S
O
W
tC1
I
:
(3.44)