第三章
3.1一元线性回归
假如说现在有一个正态分布,正态分布由mu和sigama决定,极大似然估计就是用来确定正态分布的这两个参数的
3.2多元线性回归
对线性回归方程进行化简
将
b
=
w
d
+
1
∗
1
b=w_{d+1}*1
b=wd+1∗1 因此要在向量
w
w
w后面补一个
w
d
+
1
w_{d+1}
wd+1 ,向量
x
x
x 后面补一个
1
1
1 ,化成两个向量内积的形式
f
(
x
i
)
=
(
w
1
w
2
⋯
w
d
w
d
+
1
)
(
x
i
1
x
i
2
⋮
x
i
d
1
)
f
(
x
^
i
)
=
w
^
T
x
^
i
\begin{gathered} f\left(\boldsymbol{x}_{i}\right)=\left(\begin{array}{lllll} w_{1} & w_{2} & \cdots & w_{d} & w_{d+1} \end{array}\right)\left(\begin{array}{c} x_{i 1} \\ x_{i 2} \\ \vdots \\ x_{i d} \\ 1 \end{array}\right) \\ f\left(\hat{\boldsymbol{x}}_{i}\right)=\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{i} \end{gathered}
f(xi)=(w1w2⋯wdwd+1)⎝⎜⎜⎜⎜⎜⎛xi1xi2⋮xid1⎠⎟⎟⎟⎟⎟⎞f(x^i)=w^Tx^i
将最小二乘法得到的式子向量化,便于用numpy计算
E
w
^
=
∑
i
=
1
m
(
y
i
−
w
^
T
x
^
i
)
2
=
(
y
1
−
w
^
T
x
^
1
)
2
+
(
y
2
−
w
^
T
x
^
2
)
2
+
…
+
(
y
m
−
w
^
T
x
^
m
)
2
E
w
^
=
(
y
1
−
w
^
T
x
^
1
y
2
−
w
^
T
x
^
2
⋯
y
m
−
w
^
T
x
^
m
)
(
y
1
−
w
^
T
x
^
1
y
2
−
w
^
T
x
^
2
⋮
y
m
−
w
^
T
x
^
m
)
\begin{aligned} &E_{\hat{\boldsymbol{w}}}=\sum_{i=1}^{m}\left(y_{i}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{i}\right)^{2}=\left(y_{1}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{1}\right)^{2}+\left(y_{2}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{2}\right)^{2}+\ldots+\left(y_{m}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{m}\right)^{2} \\ &E_{\hat{\boldsymbol{w}}}=\left(\begin{array}{llll} y_{1}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{1} & y_{2}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{2} & \cdots & y_{m}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{m} \end{array}\right)\left(\begin{array}{c} y_{1}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{1} \\ y_{2}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{2} \\ \vdots \\ y_{m}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{m} \end{array}\right) \end{aligned}
Ew^=i=1∑m(yi−w^Tx^i)2=(y1−w^Tx^1)2+(y2−w^Tx^2)2+…+(ym−w^Tx^m)2Ew^=(y1−w^Tx^1y2−w^Tx^2⋯ym−w^Tx^m)⎝⎜⎜⎜⎛y1−w^Tx^1y2−w^Tx^2⋮ym−w^Tx^m⎠⎟⎟⎟⎞
将后面的列向量转化一下形式
(
y
1
−
w
^
T
x
^
1
y
2
−
w
^
T
x
^
2
⋮
y
m
−
w
^
T
x
^
m
)
=
(
y
1
y
2
⋮
y
m
)
−
(
w
^
T
x
^
1
w
^
T
x
^
2
⋮
w
^
T
x
^
m
)
=
(
y
1
y
2
⋮
y
m
)
−
(
x
^
1
T
w
^
x
^
2
T
w
^
⋮
x
^
m
T
w
^
)
y
=
(
y
1
y
2
⋮
x
^
2
T
w
^
⋮
x
^
m
T
w
^
)
,
=
(
x
^
1
T
x
^
2
T
⋮
x
^
m
T
)
⋅
w
^
=
(
x
1
T
1
x
2
T
1
⋮
⋮
x
m
T
1
)
⋅
w
^
=
X
⋅
w
^
\begin{gathered} \left(\begin{array}{c} y_{1}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{1} \\ y_{2}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{2} \\ \vdots \\ y_{m}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{m} \end{array}\right)=\left(\begin{array}{c} y_{1} \\ y_{2} \\ \vdots \\ y_{m} \end{array}\right)-\left(\begin{array}{c} \hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{1} \\ \hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{2} \\ \vdots \\ \hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{m} \end{array}\right)=\left(\begin{array}{c} y_{1} \\ y_{2} \\ \vdots \\ y_{m} \end{array}\right)-\left(\begin{array}{c} \hat{\boldsymbol{x}}_{1}^{\mathrm{T}} \hat{\boldsymbol{w}} \\ \hat{\boldsymbol{x}}_{2}^{\mathrm{T}} \hat{\boldsymbol{w}} \\ \vdots \\ \hat{\boldsymbol{x}}_{m}^{\mathrm{T}} \hat{\boldsymbol{w}} \end{array}\right) \\ \boldsymbol{y}=\left(\begin{array}{c} y_{1} \\ y_{2} \\ \vdots \\ \hat{\boldsymbol{x}}_{2}^{\mathrm{T}} \hat{\boldsymbol{w}} \\ \vdots \\ \hat{\boldsymbol{x}}_{m}^{\mathrm{T}} \hat{\boldsymbol{w}} \end{array}\right), \quad=\left(\begin{array}{c} \hat{\boldsymbol{x}}_{1}^{\mathrm{T}} \\ \hat{\boldsymbol{x}}_{2}^{\mathrm{T}} \\ \vdots \\ \hat{\boldsymbol{x}}_{m}^{\mathrm{T}} \end{array}\right) \cdot \hat{\boldsymbol{w}}=\left(\begin{array}{cc} \boldsymbol{x}_{1}^{\mathrm{T}} & 1 \\ \boldsymbol{x}_{2}^{\mathrm{T}} & 1 \\ \vdots & \vdots \\ \boldsymbol{x}_{m}^{\mathrm{T}} & 1 \end{array}\right) \cdot \hat{\boldsymbol{w}}=\mathbf{X} \cdot \hat{\boldsymbol{w}} \end{gathered}
⎝⎜⎜⎜⎛y1−w^Tx^1y2−w^Tx^2⋮ym−w^Tx^m⎠⎟⎟⎟⎞=⎝⎜⎜⎜⎛y1y2⋮ym⎠⎟⎟⎟⎞−⎝⎜⎜⎜⎛w^Tx^1w^Tx^2⋮w^Tx^m⎠⎟⎟⎟⎞=⎝⎜⎜⎜⎛y1y2⋮ym⎠⎟⎟⎟⎞−⎝⎜⎜⎜⎛x^1Tw^x^2Tw^⋮x^mTw^⎠⎟⎟⎟⎞y=⎝⎜⎜⎜⎜⎜⎜⎜⎜⎛y1y2⋮x^2Tw^⋮x^mTw^⎠⎟⎟⎟⎟⎟⎟⎟⎟⎞,=⎝⎜⎜⎜⎛x^1Tx^2T⋮x^mT⎠⎟⎟⎟⎞⋅w^=⎝⎜⎜⎜⎛x1Tx2T⋮xmT11⋮1⎠⎟⎟⎟⎞⋅w^=X⋅w^
( y 1 − w ^ T x ^ 1 y 2 − w ^ T x ^ 2 ⋮ y m − w ^ T x ^ m ) = y − X w ^ E w ^ = ( y 1 − w ^ T x ^ 1 y 2 − w ^ T x ^ 2 ⋯ y m − w ^ T x ^ m ) ( y 1 − w ^ T x ^ 1 y 2 − w ^ T x ^ 2 ⋮ y m − w ^ T x ^ m ) E w ^ = ( y − X w ^ ) T ( y − X w ^ ) \begin{aligned} &\left(\begin{array}{c} y_{1}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{1} \\ y_{2}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{2} \\ \vdots \\ y_{m}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{m} \end{array}\right)=\boldsymbol{y}-\mathbf{X} \hat{\boldsymbol{w}}\\ &E_{\hat{\boldsymbol{w}}}=\left(\begin{array}{cccc} y_{1}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{1} & y_{2}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{2} & \cdots & y_{m}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{m} \end{array}\right)\left(\begin{array}{c} y_{1}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{1} \\ y_{2}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{2} \\ \vdots \\ y_{m}-\hat{\boldsymbol{w}}^{\mathrm{T}} \hat{\boldsymbol{x}}_{m} \end{array}\right)\\ &E_{\hat{\boldsymbol{w}}}=(\boldsymbol{y}-\mathbf{X} \hat{\boldsymbol{w}})^{\mathrm{T}}(\boldsymbol{y}-\mathbf{X} \hat{\boldsymbol{w}}) \end{aligned} ⎝⎜⎜⎜⎛y1−w^Tx^1y2−w^Tx^2⋮ym−w^Tx^m⎠⎟⎟⎟⎞=y−Xw^Ew^=(y1−w^Tx^1y2−w^Tx^2⋯ym−w^Tx^m)⎝⎜⎜⎜⎛y1−w^Tx^1y2−w^Tx^2⋮ym−w^Tx^m⎠⎟⎟⎟⎞Ew^=(y−Xw^)T(y−Xw^)
求最值得到 w ^ \hat{\boldsymbol{w}} w^
y为标量,对向量x的求导方法,就是将y对每一个x求偏导数后组成一个列向量
∂ f ( x ) ∂ x = [ ∂ f ( x ) ∂ x 1 ∂ f ( x ) ∂ x 2 ⋮ ∂ f ( x ) ∂ x n ] \frac{\partial f(\boldsymbol{x})}{\partial \boldsymbol{x}}=\left[\begin{array}{c}\frac{\partial f(\boldsymbol{x})}{\partial x_{1}} \\ \frac{\partial f(\boldsymbol{x})}{\partial x_{2}} \\ \vdots \\ \frac{\partial f(\boldsymbol{x})}{\partial x_{n}}\end{array}\right] ∂x∂f(x)=⎣⎢⎢⎢⎢⎡∂x1∂f(x)∂x2∂f(x)⋮∂xn∂f(x)⎦⎥⎥⎥⎥⎤ , ∂ f ( x ) ∂ x T = ( ∂ f ( x ) ∂ x 1 ∂ f ( x ) ∂ x 2 ⋯ ∂ f ( x ) ∂ x n ) \frac{\partial f(\boldsymbol{x})}{\partial \boldsymbol{x}^{\mathrm{T}}}=\left(\begin{array}{llll}\frac{\partial f(\boldsymbol{x})}{\partial x_{1}} & \frac{\partial f(\boldsymbol{x})}{\partial x_{2}} & \cdots & \frac{\partial f(\boldsymbol{x})}{\partial x_{n}}\end{array}\right) ∂xT∂f(x)=(∂x1∂f(x)∂x2∂f(x)⋯∂xn∂f(x))
上面左侧为分母布局(默认),右侧为分子布局,仅差一个转置(默认用分母布局)
几个常用的矩阵微分公式
∂
x
T
a
∂
x
=
∂
a
T
x
∂
x
=
a
,
∂
x
T
A
x
∂
x
=
(
A
+
A
T
)
x
\frac{\partial \boldsymbol{x}^{\mathrm{T}} \boldsymbol{a}}{\partial \boldsymbol{x}}=\frac{\partial \boldsymbol{a}^{\mathrm{T}} \boldsymbol{x}}{\partial \boldsymbol{x}}=\boldsymbol{a}, \frac{\partial \boldsymbol{x}^{\mathrm{T}} \mathbf{A} \boldsymbol{x}}{\partial \boldsymbol{x}}=\left(\mathbf{A}+\mathbf{A}^{\mathrm{T}}\right) \boldsymbol{x}
∂x∂xTa=∂x∂aTx=a,∂x∂xTAx=(A+AT)x
w ^ = ( X T X ) − 1 X T y \hat{\boldsymbol{w}}=\left(\mathbf{X}^{\mathrm{T}} \mathbf{X}\right)^{-1} \mathbf{X}^{\mathrm{T}} \boldsymbol{y} w^=(XTX)−1XTy
3.3对数几率回归
对数似然函数
ℓ
(
β
)
=
ln
L
(
β
)
=
∑
i
=
1
m
ln
p
(
y
i
∣
x
^
i
;
β
)
ℓ
(
β
)
=
∑
i
=
1
m
ln
(
y
i
p
1
(
x
^
i
;
β
)
+
(
1
−
y
i
)
p
0
(
x
^
i
;
β
)
)
\begin{gathered} \ell(\boldsymbol{\beta})=\ln L(\boldsymbol{\beta})=\sum_{i=1}^{m} \ln p\left(y_{i} \mid \hat{\boldsymbol{x}}_{i} ; \boldsymbol{\beta}\right) \\ \ell(\boldsymbol{\beta})=\sum_{i=1}^{m} \ln \left(y_{i} p_{1}\left(\hat{\boldsymbol{x}}_{i} ; \boldsymbol{\beta}\right)+\left(1-y_{i}\right) p_{0}\left(\hat{\boldsymbol{x}}_{i} ; \boldsymbol{\beta}\right)\right) \end{gathered}
ℓ(β)=lnL(β)=i=1∑mlnp(yi∣x^i;β)ℓ(β)=i=1∑mln(yip1(x^i;β)+(1−yi)p0(x^i;β))
概率表示为sigmoid函数形式
p
1
(
x
^
i
;
β
)
=
e
β
T
x
^
i
1
+
e
β
T
x
^
i
,
p
0
(
x
^
i
;
β
)
=
1
1
+
e
β
T
x
^
i
p_{1}\left(\hat{\boldsymbol{x}}_{i} ; \boldsymbol{\beta}\right)=\frac{e^{\boldsymbol{\beta}^{\mathrm{T}} \hat{\boldsymbol{x}}_{i}}}{1+e^{\boldsymbol{\beta}^{\mathrm{T}} \hat{\boldsymbol{x}}_{i}}}, p_{0}\left(\hat{\boldsymbol{x}}_{i} ; \boldsymbol{\beta}\right)=\frac{1}{1+e^{\boldsymbol{\beta}^{\mathrm{T}} \hat{\boldsymbol{x}}_{i}}}
p1(x^i;β)=1+eβTx^ieβTx^i,p0(x^i;β)=1+eβTx^i1
最后可得 ℓ ( β ) = ∑ i = 1 m ( y i β T x ^ i − ln ( 1 + e β T x ^ i ) ) \ell(\boldsymbol{\beta})=\sum_{i=1}^{m}\left(y_{i} \boldsymbol{\beta}^{\mathrm{T}} \hat{\boldsymbol{x}}_{i}-\ln \left(1+e^{\boldsymbol{\beta}^{\mathrm{T}} \hat{\boldsymbol{x}}_{i}}\right)\right) ℓ(β)=∑i=1m(yiβTx^i−ln(1+eβTx^i))
由于损失函数通常是以最小化为优化目标,因此可以将最大化 ℓ ( β ) \ell(\boldsymbol{\beta}) ℓ(β) 等价转化为最小化 ℓ ( β ) \ell(\boldsymbol{\beta}) ℓ(β) 的相反数 − ℓ ( β ) -\ell(\boldsymbol{\beta}) −ℓ(β)
信息论
信息论:以概率论、随机过程为基本研究工具,研究广义通信系统的整个过程。常见的应用有无损数据压缩(如ZIP文件)、有损数据压缩(如MP3和JPEG)等,本节仅引用部分精华内容。
自信息:现在有一个随机变量
X
X
X,还有他的概率密度
p
p
p,自信息就是下式
I
(
X
)
=
−
log
b
p
(
x
)
I(X)=-\log _{b} p(x)
I(X)=−logbp(x)
信息熵(自信息的期望):度量随机变量
X
X
X 的不确定性,信息熵越大越不确定
H
(
X
)
=
E
[
I
(
X
)
]
=
−
∑
x
p
(
x
)
log
b
p
(
x
)
(此处以离散型为例)
H(X)=E[I(X)]=-\sum_{x} p(x) \log _{b} p(x) \quad \text { (此处以离散型为例) }
H(X)=E[I(X)]=−x∑p(x)logbp(x) (此处以离散型为例)
相对熵(KL散度):度量两个分布的差异,其典型使用场景是用来度量理想分布
p
(
x
)
p(x)
p(x) 和模拟分布
q
(
x
)
q(x)
q(x) 之间的差异。
一般不知道
p
(
x
)
p(x)
p(x) ,我们需要让
q
(
x
)
q(x)
q(x) 离
p
(
x
)
p(x)
p(x) 越接近越好
D
K
L
(
p
∥
q
)
=
∑
x
p
(
x
)
log
b
(
p
(
x
)
q
(
x
)
)
=
∑
x
p
(
x
)
(
log
b
p
(
x
)
−
log
b
q
(
x
)
)
=
∑
x
p
(
x
)
log
b
p
(
x
)
−
∑
x
p
(
x
)
log
b
q
(
x
)
\begin{aligned} D_{K L}(p \| q) &=\sum_{x} p(x) \log _{b}\left(\frac{p(x)}{q(x)}\right) \\ &=\sum_{x} p(x)\left(\log _{b} p(x)-\log _{b} q(x)\right) \\ &=\sum_{x} p(x) \log _{b} p(x)-\sum_{x} p(x) \log _{b} q(x) \end{aligned}
DKL(p∥q)=x∑p(x)logb(q(x)p(x))=x∑p(x)(logbp(x)−logbq(x))=x∑p(x)logbp(x)−x∑p(x)logbq(x)
其中
−
∑
x
p
(
x
)
log
b
q
(
x
)
-\sum_{x} p(x) \log _{b} q(x)
−∑xp(x)logbq(x) 称为交叉熵,
∑
x
p
(
x
)
log
b
p
(
x
)
\sum_{x} p(x) \log _{b} p(x)
∑xp(x)logbp(x) 我们一般理解为常数项,因为
p
(
x
)
p(x)
p(x) 未知但固定
从机器学习三要素中“策略”的角度来说,与理想分布最接近的模拟分布即为最优分布,因此可以通过最小化相对熵这个策略来求出最优分布。
最小化相对熵,就相当于最小化交叉熵
m个样本的全体交叉熵为
∑
i
=
1
m
[
−
y
i
ln
p
1
(
x
^
i
;
β
)
−
(
1
−
y
i
)
ln
p
0
(
x
^
i
;
β
)
]
\sum_{i=1}^{m}\left[-y_{i} \ln p_{1}\left(\hat{\boldsymbol{x}}_{i} ; \boldsymbol{\beta}\right)-\left(1-y_{i}\right) \ln p_{0}\left(\hat{\boldsymbol{x}}_{i} ; \boldsymbol{\beta}\right)\right]
i=1∑m[−yilnp1(x^i;β)−(1−yi)lnp0(x^i;β)]
3.4二分类线性判别分析
让全体训练样本经过投影后:
- 异类样本的中心尽可能远
一般我们转化为内积的形式,我们一般对正负样本投影都乘以
∣
w
∣
|w|
∣w∣,因此
∣
μ
1
∣
⋅
cos
θ
2
\left|\boldsymbol{\mu}_{1}\right| \cdot \cos \theta_{2}
∣μ1∣⋅cosθ2 、
∣
μ
0
∣
⋅
cos
θ
0
\left|\boldsymbol{\mu}_{0}\right| \cdot \cos \theta_{0}
∣μ0∣⋅cosθ0 分别为正负样本的中心在
w
w
w 反向上的投影,二者的投影都乘以同一个模长,不影响求二者的差的最大值
max
∥
w
T
μ
0
−
w
T
μ
1
∥
2
2
max
∥
∣
w
∣
⋅
∣
μ
0
∣
⋅
cos
θ
0
−
∣
w
∣
⋅
∣
μ
1
∣
⋅
cos
θ
1
∥
2
2
\begin{gathered} \max \left\|\boldsymbol{w}^{\mathrm{T}} \boldsymbol{\mu}_{0}-\boldsymbol{w}^{\mathrm{T}} \boldsymbol{\mu}_{1}\right\|_{2}^{2} \\ \max \left\||\boldsymbol{w}| \cdot\left|\boldsymbol{\mu}_{0}\right| \cdot \cos \theta_{0}-|\boldsymbol{w}| \cdot\left|\boldsymbol{\mu}_{1}\right| \cdot \cos \theta_{1}\right\|_{2}^{2} \end{gathered}
max∥∥wTμ0−wTμ1∥∥22max∥∣w∣⋅∣μ0∣⋅cosθ0−∣w∣⋅∣μ1∣⋅cosθ1∥22
- 同类样本的方差(实际上并不是严格意义上的方差,因为没有除以样本总数)尽可能小
将所有点都投影到w上
min
w
T
Σ
0
w
w
T
Σ
0
w
=
w
T
(
∑
x
∈
X
0
(
x
−
μ
0
)
(
x
−
μ
0
)
T
)
w
=
∑
x
∈
X
0
(
w
T
x
−
w
T
μ
0
)
(
x
T
w
−
μ
0
T
w
)
\begin{aligned} \min \boldsymbol{w}^{\mathrm{T}} \boldsymbol{\Sigma}_{0} \boldsymbol{w} \\ \boldsymbol{w}^{\mathrm{T}} \boldsymbol{\Sigma}_{0} \boldsymbol{w} &=\boldsymbol{w}^{\mathrm{T}}\left(\sum_{\boldsymbol{x} \in X_{0}}\left(\boldsymbol{x}-\boldsymbol{\mu}_{0}\right)\left(\boldsymbol{x}-\boldsymbol{\mu}_{0}\right)^{\mathrm{T}}\right) \boldsymbol{w} \\ &=\sum_{\boldsymbol{x} \in X_{0}}\left(\boldsymbol{w}^{\mathrm{T}} \boldsymbol{x}-\boldsymbol{w}^{\mathrm{T}} \boldsymbol{\mu}_{0}\right)\left(\boldsymbol{x}^{\mathrm{T}} \boldsymbol{w}-\boldsymbol{\mu}_{0}^{\mathrm{T}} \boldsymbol{w}\right) \end{aligned}
minwTΣ0wwTΣ0w=wT(x∈X0∑(x−μ0)(x−μ0)T)w=x∈X0∑(wTx−wTμ0)(xTw−μ0Tw)
推导损失函数
max
J
=
w
T
(
μ
0
−
μ
1
)
(
μ
0
−
μ
1
)
T
w
w
T
(
Σ
0
+
Σ
1
)
w
⇓
max
J
=
w
T
S
b
w
w
T
S
w
w
\begin{gathered} \max J=\frac{\boldsymbol{w}^{\mathrm{T}}\left(\boldsymbol{\mu}_{0}-\boldsymbol{\mu}_{1}\right)\left(\boldsymbol{\mu}_{0}-\boldsymbol{\mu}_{1}\right)^{\mathrm{T}} \boldsymbol{w}}{\boldsymbol{w}^{\mathrm{T}}\left(\boldsymbol{\Sigma}_{0}+\boldsymbol{\Sigma}_{1}\right) \boldsymbol{w}} \\ \Downarrow \\ \max J=\frac{\boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{b} \boldsymbol{w}}{\boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{w} \boldsymbol{w}} \end{gathered}
maxJ=wT(Σ0+Σ1)wwT(μ0−μ1)(μ0−μ1)Tw⇓maxJ=wTSwwwTSbw
w
w
w 的大小不影响整个式子的结果,取任意值扩大或缩小相同的倍数上下会同时约掉
给定了样本后, S w S_w Sw 是个固定的常量,于是我们想要固定 w w w ,可以直接将分母固定为1,方便后续的计算
最大化转化为最小化求
w
w
w
min
w
−
w
T
S
b
w
s.t.
w
T
S
w
w
=
1
\begin{array}{cl} \min _{\boldsymbol{w}} & -\boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{b} \boldsymbol{w} \\ \text { s.t. } & \boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{w} \boldsymbol{w}=1 \end{array}
minw s.t. −wTSbwwTSww=1
解带约束的优化问题,通常用拉格朗日乘子法
只能保证最后得到的是局部极值点,但不确实是最大值还是最小值
求解
w
w
w
min
w
−
w
T
S
b
w
s.t.
w
T
S
w
w
=
1
⇔
w
T
S
w
w
−
1
=
0
\begin{array}{cl} \min _{\boldsymbol{w}} & -\boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{b} \boldsymbol{w} \\ \text { s.t. } & \boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{w} \boldsymbol{w}=1 \Leftrightarrow \boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{w} \boldsymbol{w}-1=0 \end{array}
minw s.t. −wTSbwwTSww=1⇔wTSww−1=0
由拉格朗日乘子法可得拉格朗日函数为
L
(
w
,
λ
)
=
−
w
T
S
b
w
+
λ
(
w
T
S
w
w
−
1
)
L(\boldsymbol{w}, \lambda)=-\boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{b} \boldsymbol{w}+\lambda\left(\boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{w} \boldsymbol{w}-1\right)
L(w,λ)=−wTSbw+λ(wTSww−1)
对
w
w
w 求偏导可得
∂
L
(
w
,
λ
)
∂
w
=
−
∂
(
w
T
S
b
w
)
∂
w
+
λ
∂
(
w
T
S
w
w
−
1
)
∂
w
=
−
(
S
b
+
S
b
T
)
w
+
λ
(
S
w
+
S
w
T
)
w
\begin{aligned} \frac{\partial L(\boldsymbol{w}, \lambda)}{\partial \boldsymbol{w}} &=-\frac{\partial\left(\boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{b} \boldsymbol{w}\right)}{\partial \boldsymbol{w}}+\lambda \frac{\partial\left(\boldsymbol{w}^{\mathrm{T}} \mathbf{S}_{w} \boldsymbol{w}-1\right)}{\partial \boldsymbol{w}} \\ &=-\left(\mathbf{S}_{b}+\mathbf{S}_{b}^{\mathrm{T}}\right) \boldsymbol{w}+\lambda\left(\mathbf{S}_{w}+\mathbf{S}_{w}^{\mathrm{T}}\right) \boldsymbol{w} \end{aligned}
∂w∂L(w,λ)=−∂w∂(wTSbw)+λ∂w∂(wTSww−1)=−(Sb+SbT)w+λ(Sw+SwT)w
因为
S
b
S_b
Sb 和
S
w
S_w
Sw 都是对称阵,所以
S
b
=
S
b
T
,
S
w
=
S
w
T
\mathbf{S}_{b}=\mathbf{S}_{b}^{\mathrm{T}}, \mathbf{S}_{w}=\mathbf{S}_{w}^{\mathrm{T}}
Sb=SbT,Sw=SwT
可得
∂
L
(
w
,
λ
)
∂
w
=
−
2
S
b
w
+
2
λ
S
w
w
\frac{\partial L(\boldsymbol{w}, \lambda)}{\partial \boldsymbol{w}}=-2 \mathbf{S}_{b} \boldsymbol{w}+2 \lambda \mathbf{S}_{w} \boldsymbol{w}
∂w∂L(w,λ)=−2Sbw+2λSww
令上式等于0可得
−
2
S
b
w
+
2
λ
S
w
w
=
0
S
b
w
=
λ
S
w
w
(
μ
0
−
μ
1
)
(
μ
0
−
μ
1
)
T
w
=
λ
S
w
w
\begin{aligned} &-2 \mathbf{S}_{b} \boldsymbol{w}+2 \lambda \mathbf{S}_{w} \boldsymbol{w}=0\\ &\mathbf{S}_{b} \boldsymbol{w}=\lambda \mathbf{S}_{w} \boldsymbol{w}\\ &\left(\boldsymbol{\mu}_{0}-\boldsymbol{\mu}_{1}\right)\left(\boldsymbol{\mu}_{0}-\boldsymbol{\mu}_{1}\right)^{\mathrm{T}} \boldsymbol{w}=\lambda \mathbf{S}_{w} \boldsymbol{w} \end{aligned}
−2Sbw+2λSww=0Sbw=λSww(μ0−μ1)(μ0−μ1)Tw=λSww
现在要求广义特征值
λ
\lambda
λ
若令
(
μ
0
−
μ
1
)
T
w
=
γ
\left(\boldsymbol{\mu}_{0}-\boldsymbol{\mu}_{1}\right)^{\mathrm{T}} \boldsymbol{w}=\gamma
(μ0−μ1)Tw=γ ,则
γ
(
μ
0
−
μ
1
)
=
λ
S
w
w
w
=
γ
λ
S
w
−
1
(
μ
0
−
μ
1
)
\begin{aligned} &\gamma\left(\boldsymbol{\mu}_{0}-\boldsymbol{\mu}_{1}\right)=\lambda \mathbf{S}_{w} \boldsymbol{w} \\ &\boldsymbol{w}=\frac{\gamma}{\lambda} \mathbf{S}_{w}^{-1}\left(\boldsymbol{\mu}_{0}-\boldsymbol{\mu}_{1}\right) \end{aligned}
γ(μ0−μ1)=λSwww=λγSw−1(μ0−μ1)
等式左边的
(
μ
0
−
μ
1
)
(
μ
0
−
μ
1
)
T
\left(\boldsymbol{\mu}_{0}-\boldsymbol{\mu}_{1}\right)\left(\boldsymbol{\mu}_{0}-\boldsymbol{\mu}_{1}\right)^{\mathrm{T}}
(μ0−μ1)(μ0−μ1)T 是一个行向量乘以一个列向量,得到的是一个常量,因此大小主要与
w
w
w (可调控)有关
γ \gamma γ 的值可由 w w w 调控
广义特征值
广义瑞利商
厄米矩阵
这里仅考虑实数域,A、B为对称阵,