算法描述:
Logistic Regression Algorithm |
---|
初始化
ω
0
{\omega _0}
ω0 For t = 0 , 1 , 2 , ⋯ t=0,1,2, \cdots t=0,1,2,⋯ 1.计算梯度方向: ∇ E i n ( ω t ) = 1 N ∑ n = 1 N θ ( − y n ω t T x n ) ( − y n x n ) \nabla {E_{in}}({\omega _t}) = \frac{1}{N}\sum\limits_{n = 1}^N {\theta ( - {y_n}\omega _t^T{x_n})( - {y_n}{x_n})} ∇Ein(ωt)=N1n=1∑Nθ(−ynωtTxn)(−ynxn) 2.更新: ω t + 1 ← ω t − η ∇ E i n ( ω t ) {\omega _{t + 1}} \leftarrow {\omega _t} - \eta \nabla {E_{in}}({\omega _t}) ωt+1←ωt−η∇Ein(ωt) Until ∇ E i n ( ω t + 1 ) = 0 \nabla {E_{in}}({\omega _{t + 1}}) = 0 ∇Ein(ωt+1)=0,或者足够的次数 |
这里的目标函数:
f
(
x
)
=
P
(
+
1
∣
x
)
∈
[
0
,
1
]
f(x) = P( + 1\left| x \right.) \in \left[ {0,1} \right]
f(x)=P(+1∣x)∈[0,1] ,用于二分类,则当
f
(
x
)
>
0.5
f(x) > 0.5
f(x)>0.5 ,为+1;当
f
(
x
)
<
0.5
f(x) < 0.5
f(x)<0.5 ,为-1。
计算过程:
Logistic Function:
θ ( s ) = e s 1 + e s = 1 1 + e − s \theta (s) = \frac{{{e^s}}}{{1 + {e^s}}} = \frac{1}{{1 + {e^{ - s}}}} θ(s)=1+eses=1+e−s1
图像如下,
![在这里插入图片描述](https://i-blog.csdnimg.cn/blog_migrate/5b5476edc3ccbd4ca9c55cd7429c282e.png)
该函数的特性:
- 定义域 ( − ∞ , + ∞ ) ( - \infty , + \infty ) (−∞,+∞)
- 值域 ( 0 , 1 ) (0,1) (0,1)
- 在定义域内是smooth,monotonic,sigmiod的
- θ ( s ) = 1 − θ ( − s ) \theta (s) = 1 - \theta ( - s) θ(s)=1−θ(−s)
- d θ ( s ) d s = θ ( s ) ( 1 − θ ( s ) ) \frac{{d\theta (s)}}{{ds}} = \theta (s)(1 - \theta (s)) dsdθ(s)=θ(s)(1−θ(s))
logistic函数用在逻辑回归里为,
h
(
x
)
=
1
1
+
exp
(
−
ω
T
x
)
h(x) = \frac{1}{{1 + \exp ( - {\omega ^T}x)}}
h(x)=1+exp(−ωTx)1
下面根据极大似然原理(Maximum Likelihood) 来计算逻辑回归的参数更新式。
现有目标函数如下,
f
(
x
)
=
P
(
+
1
∣
x
)
⇔
P
(
y
∣
x
)
=
{
f
(
x
)
f
o
r
y
=
+
1
1
−
f
(
x
)
f
o
r
y
=
−
1
f(x) = P( + 1\left| x \right.) \Leftrightarrow P(y\left| x \right.) = \left\{ \begin{array}{l} f(x){\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} for{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} y = + 1{\kern 1pt} \\ 1 - f(x){\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} for{\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} y = - 1 \end{array} \right.
f(x)=P(+1∣x)⇔P(y∣x)={f(x)fory=+11−f(x)fory=−1
假设现在有资料集 D = { ( x 1 , ◯ ) , ( x 2 , × ) , ⋯   , ( x N , × ) } D = \{ ({x_1},\bigcirc ),({x_2}, \times ), \cdots ,({x_N}, \times )\} D={(x1,◯),(x2,×),⋯,(xN,×)}。
则通过h产生数据集D的可能性为:
P
(
x
1
)
h
(
x
1
)
×
P
(
x
2
)
(
1
−
h
(
x
2
)
)
×
⋯
×
P
(
x
N
)
(
1
−
h
(
x
N
)
)
P({x_1})h({x_1}) \times P({x_2})(1 - h({x_2})) \times \cdots \times P({x_N})(1 - h({x_N}))
P(x1)h(x1)×P(x2)(1−h(x2))×⋯×P(xN)(1−h(xN))
通常由目标函数f产生数据集D的概率是很大的 (极大似然的思想),当
h
≈
f
h \approx f
h≈f时,由h产生D的概率也是非常大的,即,
g
≈
arg
max
h
l
i
k
e
l
i
h
o
o
d
(
h
)
g \approx \mathop {\arg \max }\limits_h {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} likelihood(h)
g≈hargmaxlikelihood(h)
这里
h
(
x
)
=
θ
(
ω
T
x
)
h(x) = \theta ({\omega ^T}x)
h(x)=θ(ωTx),又有
1
−
h
(
x
)
=
h
(
−
x
)
1 - h(x) = h( - x)
1−h(x)=h(−x),所以,
l
i
k
e
l
i
h
o
o
d
(
h
)
=
P
(
x
1
)
h
(
x
1
)
×
P
(
x
2
)
(
1
−
h
(
x
2
)
)
×
⋯
×
P
(
x
N
)
(
1
−
h
(
x
N
)
)
=
P
(
x
1
)
h
(
x
1
)
×
P
(
x
2
)
h
(
−
x
2
)
×
⋯
×
P
(
x
N
)
h
(
−
x
N
)
=
P
(
x
1
)
h
(
y
1
x
1
)
×
P
(
x
2
)
h
(
y
2
x
2
)
×
⋯
×
P
(
x
N
)
h
(
y
N
x
N
)
\begin{array}{l} likelihood(h) = P({x_1})h({x_1}) \times P({x_2})(1 - h({x_2})) \times \\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} \cdots \times P({x_N})(1 - h({x_N}))\\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} = P({x_1})h({x_1}) \times P({x_2})h( - {x_2}) \times \\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} \cdots \times P({x_N})h( - {x_N})\\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} = P({x_1})h({y_1}{x_1}) \times P({x_2})h({y_2}{x_2}) \times \\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} \cdots \times P({x_N})h({y_N}{x_N}) \end{array}
likelihood(h)=P(x1)h(x1)×P(x2)(1−h(x2))×⋯×P(xN)(1−h(xN))=P(x1)h(x1)×P(x2)h(−x2)×⋯×P(xN)h(−xN)=P(x1)h(y1x1)×P(x2)h(y2x2)×⋯×P(xN)h(yNxN)
对于每个不同的h而言,
P
(
x
i
)
P({x_i})
P(xi)都是不变的,那么就有,
l
i
k
e
l
i
h
o
o
d
(
h
)
∝
∏
n
=
1
N
h
(
y
n
x
n
)
likelihood(h) \propto \prod\limits_{n = 1}^N {h({y_n}{x_n})}
likelihood(h)∝n=1∏Nh(ynxn)
用
ω
\omega
ω表示h,有,
max
ω
l
i
k
e
l
i
h
o
o
d
(
h
)
∝
∏
n
=
1
N
θ
(
y
n
ω
T
x
n
)
\mathop {\max }\limits_\omega likelihood(h) \propto \prod\limits_{n = 1}^N {\theta ({y_n}{\omega ^T}{x_n})}
ωmaxlikelihood(h)∝n=1∏Nθ(ynωTxn)
取对数 (变连乘为连加),加负号 (变最大化为最小化),再取均值,有,
min
ω
1
N
∑
n
=
1
N
−
ln
θ
(
y
n
ω
T
x
n
)
=
min
ω
1
N
∑
n
=
1
N
ln
(
1
+
exp
(
−
y
n
ω
T
x
n
)
)
=
min
ω
1
N
∑
n
=
1
N
e
r
r
(
ω
,
x
n
,
y
n
)
⎵
E
i
n
(
ω
)
\begin{array}{l} \mathop {\min }\limits_\omega \frac{1}{N}\sum\limits_{n = 1}^N { - \ln \theta ({y_n}{\omega ^T}{x_n})} \\ = \mathop {\min }\limits_\omega \frac{1}{N}\sum\limits_{n = 1}^N {\ln (1 + \exp ( - {y_n}{\omega ^T}{x_n}))} \\ {\kern 1pt} = \mathop {\min }\limits_\omega \frac{1}{N}\underbrace {\sum\limits_{n = 1}^N {err(\omega ,{x_n},{y_n})} }_{{E_{in}}(\omega )} \end{array}
ωminN1n=1∑N−lnθ(ynωTxn)=ωminN1n=1∑Nln(1+exp(−ynωTxn))=ωminN1Ein(ω)
n=1∑Nerr(ω,xn,yn)
上式,就是逻辑回归里的误差衡量方式——交叉熵误差(Cross-Entropy Error),即,
e
r
r
(
ω
,
x
,
y
)
=
ln
(
1
+
exp
(
−
y
ω
x
)
err(\omega ,x,y) = \ln (1 + \exp ( - y\omega x)
err(ω,x,y)=ln(1+exp(−yωx)
根据凸函数的最小化原理,令
∇
E
i
n
(
ω
)
=
0
\nabla {E_{in}}(\omega ) = 0
∇Ein(ω)=0,下面计算梯度,
E
i
n
(
ω
)
=
1
N
∑
n
=
1
N
ln
(
1
+
exp
(
−
y
n
ω
T
x
n
⏞
◯
)
⎵
Δ
)
\begin{array}{l} {E_{in}}(\omega ) = \frac{1}{N}\sum\limits_{n = 1}^N {\ln (\underbrace {1 + \exp (\overbrace { - {y_n}{\omega ^T}{x_n}}^\bigcirc )}_\Delta )} \\ {\kern 1pt} \end{array}
Ein(ω)=N1n=1∑Nln(Δ
1+exp(−ynωTxn
◯))
∂ E i n ( ω ) ∂ ω i = 1 N ∑ n = 1 N ( ∂ ln ( Δ ) ∂ Δ ) ( ∂ ( 1 + exp ( ◯ ) ) ∂ ◯ ) ( ∂ − y n ω T x n ∂ ω i ) = 1 N ∑ n = 1 N ( 1 Δ ) ( exp ( ◯ ) ) ( − y n x n , i ) = 1 N ∑ n = 1 N ( exp ( ◯ ) 1 + exp ( ◯ ) ) ( − y n x n , i ) = 1 N ∑ n = 1 N θ ◯ ( − y n x n , i ) \begin{array}{l} \frac{{\partial {E_{in}}(\omega )}}{{\partial {\omega _i}}} = \frac{1}{N}\sum\limits_{n = 1}^N {(\frac{{\partial \ln (\Delta )}}{{\partial \Delta }})} (\frac{{\partial (1 + \exp (\bigcirc ))}}{{\partial \bigcirc }})(\frac{{\partial - {y_n}{\omega ^T}{x_n}}}{{\partial {\omega _i}}})\\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} = \frac{1}{N}\sum\limits_{n = 1}^N {(\frac{1}{\Delta })} (\exp (\bigcirc ))( - {y_n}{x_{n,i}})\\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} = \frac{1}{N}\sum\limits_{n = 1}^N {(\frac{{\exp (\bigcirc )}}{{1{\rm{ + }}\exp (\bigcirc )}})} ( - {y_n}{x_{n,i}})\\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} = \frac{1}{N}\sum\limits_{n = 1}^N {\theta \bigcirc } ( - {y_n}{x_{n,i}}) \end{array} ∂ωi∂Ein(ω)=N1n=1∑N(∂Δ∂ln(Δ))(∂◯∂(1+exp(◯)))(∂ωi∂−ynωTxn)=N1n=1∑N(Δ1)(exp(◯))(−ynxn,i)=N1n=1∑N(1+exp(◯)exp(◯))(−ynxn,i)=N1n=1∑Nθ◯(−ynxn,i)
即,
∇
E
i
n
(
ω
)
=
1
N
∑
n
=
1
N
θ
(
−
y
n
ω
T
x
n
)
(
−
y
n
x
n
)
=
0
\nabla {E_{in}}(\omega ) = \frac{1}{N}\sum\limits_{n = 1}^N {\theta ( - {y_n}{\omega ^T}{x_n})( - {y_n}{x_n})} = 0
∇Ein(ω)=N1n=1∑Nθ(−ynωTxn)(−ynxn)=0
上式不存在闭式解(closed-form solution),因为,把这里的
θ
(
⋅
)
\theta ( \cdot )
θ(⋅)看作是
−
y
n
x
n
- {y_n}{x_n}
−ynxn的权重,则整个梯度式子可看作是以
θ
(
⋅
)
\theta ( \cdot )
θ(⋅)为权重的关于
−
y
n
x
n
- {y_n}{x_n}
−ynxn的加权平均,所以只有当所有的
θ
(
⋅
)
=
0
\theta ( \cdot ) = 0
θ(⋅)=0成立时,
∇
E
i
n
(
ω
)
=
0
\nabla {E_{in}}(\omega ) = 0
∇Ein(ω)=0。
1.所有
θ
(
⋅
)
=
0
\theta ( \cdot ) = 0
θ(⋅)=0 ,当且仅当
y
n
ω
T
x
n
≫
0
{y_n}{\omega ^T}{x_n} \gg 0
ynωTxn≫0 ,即该数据集线性可分,一旦数据集线性不可分,则上述梯度就不可能为0
2.权重 :
θ
(
⋅
)
=
0
\theta ( \cdot ) = 0
θ(⋅)=0是关于
ω
\omega
ω的一个非线性方程,不容易得出闭式解
所以,这里的参数更新采用的是迭代优化解(Iterative Optimization),用梯度下降法求解函数的最小化问题,
ω
t
+
1
←
ω
t
−
η
∇
E
i
n
(
ω
t
)
{\omega _{t + 1}} \leftarrow {\omega _t} - \eta \nabla {E_{in}}({\omega _t})
ωt+1←ωt−η∇Ein(ωt)
实际应用:
数据特征集D为:
x
=
[
1
1
⋯
1
x
11
x
21
⋯
x
n
1
x
12
x
22
⋯
x
n
2
⋮
⋮
⋮
⋮
x
1
d
x
2
d
⋯
x
n
d
]
⎵
(
d
+
1
)
×
N
x = \underbrace {\left[ {\begin{array}{} 1&1& \cdots &1\\ {{x_{11}}}&{{x_{21}}}& \cdots &{x{}_{n1}}\\ {{x_{12}}}&{{x_{22}}}& \cdots &{x{}_{n2}}\\ \vdots & \vdots & \vdots & \vdots \\ {{x_{1d}}}&{x{}_{2d}}& \cdots &{x{}_{nd}} \end{array}} \right]}_{(d + 1) \times N}
x=(d+1)×N
⎣⎢⎢⎢⎢⎢⎡1x11x12⋮x1d1x21x22⋮x2d⋯⋯⋯⋮⋯1xn1xn2⋮xnd⎦⎥⎥⎥⎥⎥⎤
对应的标签集为:
y = [ y 1 y 2 ⋮ y n ] ⎵ N × 1 y = \underbrace {\left[ {\begin{array}{} {{y_1}}\\ {{y_2}}\\ \vdots \\ {{y_n}} \end{array}} \right]}_{N \times 1} y=N×1 ⎣⎢⎢⎢⎡y1y2⋮yn⎦⎥⎥⎥⎤
则梯度的计算如下:
A = θ ( − y n . ∗ ( ω T x n ⏞ 1 × N ) ) ⎵ 1 × N b = − y n . ∗ x n ⎵ ( d + 1 ) × N \begin{array}{l} A = \underbrace {\theta ( - {y_n}. * (\overbrace {{\omega ^T}{x_n}}^{1 \times N}))}_{1 \times N}\\ b = \underbrace { - {y_n}. * {x_n}}_{(d + 1) \times N} \end{array} A=1×N θ(−yn.∗(ωTxn 1×N))b=(d+1)×N −yn.∗xn
∇ E i n ( ω ) = A 1 ⎵ ( 常 数 ) b 1 ⎵ ( d + 1 ) × 1 + A 2 b 2 + ⋯ + A N b N = b ⎵ ( d + 1 ) × N [ A 1 A 2 ⋮ A N ] ⎵ N × 1 \begin{array}{l} \nabla {E_{in}}(\omega ) = \underbrace {{A_1}}_{(常数)}\underbrace {{b_1}}_{(d + 1) \times 1} + {A_2}{b_2} + \cdots + {A_N}{b_N}\\ {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} {\kern 1pt} = \underbrace b_{(d + 1) \times N}\underbrace {\left[ {\begin{array}{} {{A_1}}\\ {{A_2}}\\ \vdots \\ {{A_N}} \end{array}} \right]}_{N \times 1} \end{array} ∇Ein(ω)=(常数) A1(d+1)×1 b1+A2b2+⋯+ANbN=(d+1)×N bN×1 ⎣⎢⎢⎢⎡A1A2⋮AN⎦⎥⎥⎥⎤
实际应用中,一般用线性回归求初值 ,然后再用PLA/pocket/logistic regression等方法,一般logistic regression效果要好于pocket。
随机梯度(Stochastic Gradient Descent, SGD)的使用:
以上计算的梯度的时候,是计算了在所有点处的梯度和然后再平均,这里的平均的概念可以用随机的一个梯度值来近似代替,即,
ω
t
+
1
←
ω
t
+
η
θ
(
−
y
n
ω
t
T
x
n
)
(
y
n
x
n
)
⎵
−
∇
e
r
r
(
ω
t
,
x
n
,
y
n
)
{\omega _{t + 1}} \leftarrow {\omega _t} + \eta \underbrace {\theta ( - {y_n}\omega _t^T{x_n})({y_n}{x_n})}_{ - \nabla err({\omega _t},{x_n},{y_n})}
ωt+1←ωt+η−∇err(ωt,xn,yn)
θ(−ynωtTxn)(ynxn)
随机梯度的使用体现了一个在线学习思想,即每来一个数据,就可以进行一次参数更新。
Pros: 计算代价低,适合数据量大以及在线学习的场景
Cons: 不稳定。