Kernel的概念
Kernel = Transform + inner Product
在SVM的对偶问题中,
q
n
,
m
=
y
n
y
m
z
n
T
z
m
{q_{n,m}} = {y_n}{y_m}z_n^T{z_m}
qn,m=ynymznTzm在计算内积
z
n
T
z
m
(
ϕ
(
x
n
)
T
ϕ
(
x
m
)
)
z_n^T{z_m}(\phi {({x_n})^T}\phi ({x_m}))
znTzm(ϕ(xn)Tϕ(xm)) 时没有真正移除对
d
~
\tilde d
d~的依赖,这里以Kernel的方式进行计算,以此来替代在z空间里做内积,可以很明显的提高计算效率。常用的kernel有多项式kernel(Polynomial Kernel)和高斯kernel(Gaussian Kernel)。
多项式kernel(Polynomial Kernel)
考虑一个2次的多项式转换(2nd order polynomial transform),即,
ϕ
2
(
x
)
=
(
1
,
x
1
,
⋯
 
,
x
d
,
x
1
2
,
x
1
x
2
,
⋯
 
,
x
1
x
d
,
x
2
x
1
,
x
2
2
,
⋯
 
,
x
2
x
d
,
⋯
 
,
x
d
2
)
{\phi _2}(x) = (1,{x_1}, \cdots ,{x_d},x_1^2,{x_1}{x_2}, \cdots ,{x_1}{x_d},{x_2}{x_1},x_2^2, \cdots ,{x_2}{x_d}, \cdots ,x_d^2)
ϕ2(x)=(1,x1,⋯,xd,x12,x1x2,⋯,x1xd,x2x1,x22,⋯,x2xd,⋯,xd2)
,有以下的转换,
ϕ 2 ( x ) T ϕ 2 ( x ′ ) = 1 + ∑ i = 1 d x i x i ′ + ∑ i = 1 d ∑ j = 1 d x i x j x i ′ x j ′                      = 1 + ∑ i = 1 d x i x i ′ + ∑ i = 1 d x i x i ′ ∑ j = 1 d x j x j ′                      = 1 + x T x ′ + ( x T x ′ ) ( x T x ′ ) \begin{array}{l} {\phi _2}{(x)^T}{\phi _2}({x^{'}}) = 1 + \sum\limits_{i = 1}^d {{x_i}x_i^{'}} + \sum\limits_{i = 1}^d {\sum\limits_{j = 1}^d {{x_i}{x_j}{x_i}^{'}{x_j}^{'}} } \\ \\ \;\;\;\;\;\;\;\;\;\; = 1 + \sum\limits_{i = 1}^d {{x_i}x_i^{'}} + \sum\limits_{i = 1}^d {{x_i}{x_i}^{'}\sum\limits_{j = 1}^d {{x_j}{x_j}^{'}} } \\ \\ \;\;\;\;\;\;\;\;\;\; = 1 + {x^T}{x^{'}} + ({x^T}{x^{'}})({x^T}{x^{'}})\\ \end{array} ϕ2(x)Tϕ2(x′)=1+i=1∑dxixi′+i=1∑dj=1∑dxixjxi′xj′=1+i=1∑dxixi′+i=1∑dxixi′j=1∑dxjxj′=1+xTx′+(xTx′)(xTx′)
对于二次多项式转换所对应的kernel为,
K
Φ
2
(
x
,
x
′
)
=
1
+
(
x
T
x
′
)
+
(
x
T
x
′
)
2
{K_{{\Phi _2}}}(x,x') = 1 + ({x^T}x') + {({x^T}x')^2}
KΦ2(x,x′)=1+(xTx′)+(xTx′)2
再一般点,
K
2
(
x
,
x
′
)
=
(
1
+
γ
x
T
x
′
)
2
,
γ
>
0
{K_2}(x,x') = {(1 + \gamma {x^T}x')^2},\gamma > 0
K2(x,x′)=(1+γxTx′)2,γ>0
再一般点,即可得到多项式Kernel(Polynomial Kernel),
K Q ( x , x ′ ) = ( ζ + γ x T x ′ ) Q , γ > 0 , ζ ≥ 0 {K_Q}(x,x') = {(\zeta + \gamma {x^T}x')^Q},\gamma > 0,\zeta \ge 0 KQ(x,x′)=(ζ+γxTx′)Q,γ>0,ζ≥0
这里的参数 ( ς , γ ) (\varsigma ,\gamma ) (ς,γ)影响SVM边界的属性。
常用的多项式Kernel有Linear Kernel,即, K 1 = x T x ′ {K_1} = {x^T}x' K1=xTx′。
Gaussian Kernel(Radial Basis Function kernel , RBF kernel)
高斯核函数实际上表示的是一个无限维度的转换,即表示的是以下转换,
Φ
(
x
)
=
exp
(
−
x
2
)
(
1
,
2
1
!
x
,
2
2
2
!
x
2
,
⋯
 
)
\Phi (x) = \exp ( - {x^2})(1,\sqrt {\frac{2}{{1!}}} x,\sqrt {\frac{{{2^2}}}{{2!}}} {x^2}, \cdots )
Φ(x)=exp(−x2)(1,1!2x,2!22x2,⋯)
即,
K ( x , x ′ ) = exp ( − ( x − x ′ ) 2 )                = exp ( − x 2 ) exp ( − x ′ 2 ) exp ( 2 x x ′ )              = T a y l o r exp ( − x 2 ) exp ( − x ′ 2 ) ( ∑ i = 0 ∞ ( 2 x x ′ ) i i ! )                = ∑ i = 0 ∞ ( exp ( − x 2 ) exp ( − x ′ 2 ) 2 i i 2 i i x i x ′ i )                = Φ ( x ) T Φ ( x ′ ) \begin{array}{l} K(x,x') = \exp ( - {(x - x')^2})\\ \\ \;\;\;\;\;\;\; = \exp ( - {x^2})\exp ( - {{x'}^2})\exp (2xx')\\ \\ \;\;\;\;\;\;\mathop = \limits^{Taylor} \exp ( - {x^2})\exp ( - {{x'}^2})(\sum\limits_{i = 0}^\infty {\frac{{{{(2xx')}^i}}}{{i!}}} )\\ \\ \;\;\;\;\;\;\; = \sum\limits_{i = 0}^\infty {(\exp ( - {x^2})\exp ( - {{x'}^2})\sqrt {\frac{{{2^i}}}{{i}}} \sqrt {\frac{{{2^i}}}{{i}}} {x^i}{{x'}^i})} \\ \\ \;\;\;\;\;\;\; = \Phi {(x)^T}\Phi (x') \end{array} K(x,x′)=exp(−(x−x′)2)=exp(−x2)exp(−x′2)exp(2xx′)=Taylorexp(−x2)exp(−x′2)(i=0∑∞i!(2xx′)i)=i=0∑∞(exp(−x2)exp(−x′2)i2ii2ixix′i)=Φ(x)TΦ(x′)
更一般的,高斯核的形式为,
K ( x , x ′ ) = exp ( − γ ∥ x − x ′ ∥ 2 ) , γ > 0 K(x,x') = \exp ( - \gamma {\left\| {x - x'} \right\|^2}),\gamma > 0 K(x,x′)=exp(−γ∥x−x′∥2),γ>0
高斯核用在SVM里的形式为,
g S V M ( x ) = s i g n ( ∑ S V α n y n K ( x n , x ) + b )              = s i g n ( ∑ S V α n y n exp ( − γ ∥ x − x n ∥ 2 ) + b ) \begin{array}{l} {g_{SVM}}(x) = sign(\sum\limits_{SV} {{\alpha _n}{y_n}K({x_n},x) + b} )\\ \\ \;\;\;\;\;\; = sign(\sum\limits_{SV} {{\alpha _n}{y_n}\exp ( - \gamma {{\left\| {x - {x_n}} \right\|}^2}) + b} ) \end{array} gSVM(x)=sign(SV∑αnynK(xn,x)+b)=sign(SV∑αnynexp(−γ∥x−xn∥2)+b)
由上式,可看出,sign里面的式子实际上是许多个高斯函数的线性组合,这些高斯函数是以支持向量(SVs)为均值,即中心在SVs处的高斯函数,所以Gaussian Kernel又叫做Radial Basis Function(RBF) Kernel。
对于高斯核参数 γ \gamma γ的选取, γ \gamma γ越大,所得到的SVM的边界就越复杂,越容易过拟合。
Kernel SVM
在SVM中,总共有3处使用到了Kernel,
- q n , m = y n y m z n T z m = y n y m K ( x n , x m ) {q_{n,m}} = {y_n}{y_m}z_n^T{z_m} = {y_n}{y_m}K({x_n},{x_m}) qn,m=ynymznTzm=ynymK(xn,xm)
- 计算b(from SV ( x s , y s ) ({x_s},{y_s}) (xs,ys)), b = y s − w T z s = y s − ( ∑ n = 1 N α n y n z n ) T z s = y s − ∑ n = 1 N α n y n ( K ( x n , x s ) ) b = {y_s} - {w^T}{z_s} = {y_s} - {(\sum\limits_{n = 1}^N {{\alpha _n}{y_n}} {z_n})^T}{z_s} = {y_s} - \sum\limits_{n = 1}^N {{\alpha _n}{y_n}} (K({x_n},{x_s})) b=ys−wTzs=ys−(n=1∑Nαnynzn)Tzs=ys−n=1∑Nαnyn(K(xn,xs))
- 得出结果 g S V M {g_{SVM}} gSVM, g S V M ( x ) = s i g n ( w T Φ ( x ) + b ) = s i g n ( ∑ n = 1 N α n y n K ( x n , x ) + b ) {g_{SVM}}(x) = sign({w^T}\Phi (x) + b) = sign(\sum\limits_{n = 1}^N {{\alpha _n}{y_n}K({x_n},x) + b} ) gSVM(x)=sign(wTΦ(x)+b)=sign(n=1∑NαnynK(xn,x)+b)
Kernel Hard-Margin SVM Algorithm |
---|
1.
q
n
,
m
=
y
n
y
m
K
(
x
n
,
x
m
)
;
p
=
−
1
N
;
(
A
,
c
)
  
{q_{n,m}} = {y_n}{y_m}K({x_n},{x_m});p = - {1_N};(A,c)\;
qn,m=ynymK(xn,xm);p=−1N;(A,c) 2. α ← Q P ( Q D , p , A , c ) \alpha \leftarrow QP({Q_D},p,A,c) α←QP(QD,p,A,c) 3. b ← ( y s − ∑ S V α n y n K ( x n , x s ) ) , S V ( x s , y s ) b \leftarrow ({y_s} - \sum\limits_{SV} {{\alpha _n}{y_n}K({x_n},{x_s})} ),SV({x_s},{y_s}) b←(ys−SV∑αnynK(xn,xs)),SV(xs,ys) 4.返回参数 S V s & α n & b SVs\& {\alpha _n}\& b SVs&αn&b,对于一个新的数据x,其预测值为, g S V M ( x ) = s i g n ( ∑ S V α n y n K ( x n , x ) + b ) {g_{SVM}}(x) = sign(\sum\limits_{SV} {{\alpha _n}{y_n}K({x_n},x)} + b) gSVM(x)=sign(SV∑αnynK(xn,x)+b) |