标题:Multiview Subspace Clustering by an Enhanced Tensor Nuclear Norm (Schatten p-norm)
期刊 + 时间 + 有无源代码: IEEE TRANSACTIONS ON CYBERNETICS + 2021 + 有
作者: Wei Xia, Xiangdong Zhang, Quanxue Gao , Xiaochuang Shu, Jungong Han , and Xinbo Gao
方法名及缩写: weighted tensor Schatten p-norm minimization (WTSNM) based on t-SVD,
算法框架:
流程图 :
主要创新点: 提出了Schatten p-norm
动机: t-SVD中使用的张量核范数最小化(TNNM)对每个奇异值均等正则化,这在矩阵完成和系数矩阵学习中没有意义。在这种情况下,奇异值代表不同的视角,应该区别对待。
目标函数:
WTSNM
argmin
X
1
2
∥
X
−
A
∥
F
2
+
τ
∥
X
∥
ω
,
S
p
p
\operatorname*{argmin}_{\mathcal{X}}\frac12\|\mathcal{X}-\mathcal{A}\|_{F}^2+\tau\|\mathcal{X}\|_{\omega,S_p}^p
Xargmin21∥X−A∥F2+τ∥X∥ω,Spp
MVC based on WTSNM
min
Z
,
E
(
ν
)
,
F
∥
Z
∥
ω
,
S
p
p
+
λ
∥
E
∥
2
,
1
+
2
α
t
r
(
F
T
L
Z
^
F
)
s
.
t
.
X
(
ν
)
=
X
(
ν
)
Z
(
ν
)
+
E
(
ν
)
,
v
=
1
,
2
,
…
,
m
\begin{aligned}\min_{\mathcal{Z},\mathbf{E}^{(\nu)},\mathbf{F}}&\|\mathcal{Z}\|_{\boldsymbol{\omega},\boldsymbol{S}_p}^p+\lambda\|\mathbf{E}\|_{2,1}+2\alpha tr(\mathbf{F}^T\mathbf{L}_{\hat{\mathbf{Z}}}\mathbf{F})\\\mathrm{s.t.}&\mathbf{X}^{(\nu)}=\mathbf{X}^{(\nu)}\mathbf{Z}^{(\nu)}+\mathbf{E}^{(\nu)},v=1,2,\ldots,m\end{aligned}
Z,E(ν),Fmins.t.∥Z∥ω,Spp+λ∥E∥2,1+2αtr(FTLZ^F)X(ν)=X(ν)Z(ν)+E(ν),v=1,2,…,m
优化步骤:
Updata
Z
(
v
)
Z^{(v)}
Z(v): (要注意对
T
r
(
F
T
L
Z
F
)
Tr(F^TL_{Z}F)
Tr(FTLZF)项的处理,查阅原文)
Z
t
(
ν
)
∗
=
(
μ
(
X
(
ν
)
)
T
X
(
ν
)
+
ρ
I
)
−
1
(
μ
(
X
(
ν
)
)
T
X
(
ν
)
+
(
X
(
ν
)
)
T
Y
(
ν
)
+
ρ
J
(
ν
)
−
μ
(
X
(
ν
)
)
T
E
(
ν
)
−
W
(
ν
)
−
α
2
m
(
P
⊙
s
i
g
n
(
Z
t
−
1
(
ν
)
)
+
P
T
⊙
s
i
g
n
(
Z
t
−
1
(
ν
)
)
T
)
)
\begin{aligned} \mathbf{Z}_{t}^{(\nu)^{*}}& =\left(\mu(\mathbf{X}^{(\nu)})^{T}\mathbf{X}^{(\nu)}+\rho\mathbf{I}\right)^{-1} \\ &\left(\mu(\mathbf{X}^{(\nu)})^{T}\mathbf{X}^{(\nu)}+\left(\mathbf{X}^{(\nu)}\right)^{T}\mathbf{Y}^{(\nu)}+\rho J^{(\nu)}\right. \\ &-\mu(\mathbf{X}^{(\nu)})^{T}\mathbf{E}^{(\nu)}-\mathbf{W}^{(\nu)} \\ &-\frac{\alpha}{2m}\Big(\mathbf{P}\odot sign(\mathbf{Z}_{t-1}^{(\nu)})+\mathbf{P}^{T}\odot sign(\mathbf{Z}_{t-1}^{(\nu)})^{T}\Big)\Big) \end{aligned}
Zt(ν)∗=(μ(X(ν))TX(ν)+ρI)−1(μ(X(ν))TX(ν)+(X(ν))TY(ν)+ρJ(ν)−μ(X(ν))TE(ν)−W(ν)−2mα(P⊙sign(Zt−1(ν))+PT⊙sign(Zt−1(ν))T))
Update
E
E
E : Update
F
F
F : 没什么特殊
Update
J
\mathcal {J}
J : weighted Schatten p-norm的求解范式
J
∗
=
arg
min
J
∥
J
∥
ω
,
S
p
p
+
⟨
Q
,
Z
−
J
⟩
+
ρ
2
∥
Z
−
J
∥
F
2
=
arg
min
J
∥
J
∥
ω
,
S
p
p
+
ρ
2
∥
Z
−
J
+
Q
ρ
∥
F
2
\begin{gathered} \mathcal{J}^{*} =\arg\min_{\mathcal J}\|{\mathcal J}\|_{\omega,S_{p}}^{p}+\langle{\mathcal Q},{\mathcal Z}-{\mathcal J}\rangle+\frac{\rho}{2}\|{\mathcal Z}-{\mathcal J}\|_{F}^{2} \\ =\arg\min_{\mathcal{J}}\|\mathcal{J}\|_{\omega,S_{p}}^{p}+\frac{\rho}{2}\left\|\mathcal{Z}-\mathcal{J}+\frac{\mathcal{Q}}{\rho}\right\|_{F}^{2} \end{gathered}
J∗=argJmin∥J∥ω,Spp+⟨Q,Z−J⟩+2ρ∥Z−J∥F2=argJmin∥J∥ω,Spp+2ρ
Z−J+ρQ
F2