[Resolved] 偏导问题

此问题来源于模糊C均值聚类的推导过程

---

符号定义：$\left \{ x_{i}, i=1, 2, \cdots , n \right \}$是n个样本组成的样本集合，$c$为预定的类别数目，$\mu_{i}, i =1,2,\cdots , c \quad$每个聚类的中心，$\mu_{j}\left(x_{i} \right)$是第$i$个样本对于第$j$类的隶属度函数，且其满足如下关系式：

$$\sum_{j=1}^{c}\mu_{j} \left( x_{i} \right)=1 \quad i=1, 2, \cdots,c \quad\quad\text{(1)}$$

用隶属度函数定义的聚类损失函数可以写为：

$$J_{f}=\sum_{j=1}^c \sum_{i=1}^{n} \left[\mu_{j}(x_{i}) \right]^{b} \left\| x_{i} - \mu_{j} \right\| ^{2}\quad\quad\text{(2)}$$
其中， $b>1$是一个可以控制聚类结果的隶属度程度的常数。

令$J_f$对$\mu_{j}\left(x_i\right)$求偏导，并令偏导结果为0，可得：

$$\mu_{j}\left ( x_{i} \right) = \frac{\left [ 1/\left \| x_{i} - \mu_{j} \right \| ^{2} \right ]^{1/(b-1)}} {\displaystyle\sum_{k=1}^{c} \left [ 1/\left \| x_{i} - \mu_{k} \right \| ^{2} \right ]^{1/(b-1)}} \quad i=1, 2, \cdots, n \quad j=1, 2, \cdots, c \quad\quad\text{(3)}$$

The Question is:

How to get the result $(3)$ from $(2)$ ?

Solution: Thanks to Mr.H

记 $g=\displaystyle\sum_{j=1}^{c}\mu_j\left(x_i\right)-1$，则原问题可描述为：在 $g=0$ 的条件下， 求$J_f$最小。
使

$$J_f-\lambda g$$ 对 $\mu_j\left(x_i\right),\quad j=1, 2, \cdots, c$ 求偏导。

得：

$$\left \{ \begin{aligned} b \cdot \left [ \mu_1\left(x_i\right) \right ]^{b-1} \cdot \| x_i-\mu_1 \| ^2 &= \lambda \\ b \cdot \left [ \mu_2\left(x_i\right) \right ]^{b-1} \cdot \| x_i-\mu_2 \| ^2 &= \lambda \\ \cdots \\ b \cdot \left [ \mu_c\left(x_i\right) \right ]^{b-1} \cdot \| x_i-\mu_c \| ^2 &= \lambda \\ \sum_{j=1}^c\mu_j\left( x_i\right) - 1&=0 \end{aligned} \right. \qquad , i=1, 2, \cdots, n$$

即：

$$\frac{\lambda}{b}=\left [ \mu_j\left(x_i\right)\right ]^{b-1}\cdot \| x_i - \mu_j\|^2 \qquad\qquad ,\left \{ \begin{aligned} i=1, 2, \cdots, n \\ j=1, 2, \cdots, c \end{aligned} \right. \qquad(1)$$
$$\left(\frac{\lambda}{b}\right)^{1/(b-1)}= \mu_j\left(x_i\right)\cdot \left[\| x_i - \mu_j\|^2\right]^{1/(b-1)} \qquad\qquad ,\left \{ \begin{aligned} i=1, 2, \cdots, n \\ j=1, 2, \cdots, c \end{aligned} \right. \qquad(2)$$

$$\mu_j\left(x_i\right)=\left(\frac{\lambda}{b}\right)^{1/(b-1)}\cdot \left[1/\|x_i-\mu_j\|^2\right]^{1/(b-1)} \qquad\qquad ,\left \{ \begin{aligned} i=1, 2, \cdots, n \\ j=1, 2, \cdots, c \end{aligned} \right. \qquad(3)$$

而后：
$\because$

$$\displaystyle\sum_{j=1}^{c}\mu_j\left(x_i\right)=1 \qquad ,i=1, 2, \cdots, c$$

$\therefore$ 式$(3)$变为:

$$\left(\frac{\lambda}{b}\right)^{1/(b-1)} \sum_{j=1}^c \left[1/\|x_i-\mu_j\|^2\right]^{1/(b-1)}=1$$

$$\left(\frac{\lambda}{b}\right)^{1/(b-1)}=\frac{1}{\displaystyle\sum_{j=1}^c\left[1/\|x_i-\mu_j\|^2\right]^{1/(b-1)}} \qquad\qquad(4)$$

$(4)\rightarrow(2)$，得：

$$\frac{1}{\displaystyle\sum_{j=1}^c\left[1/\|x_i-\mu_j\|^2\right]^{1/(b-1)}} =\mu_j\left(x_i\right)\cdot\left[\|x_i-\mu_j\|^2\right]^{1/(b-1)} \qquad\qquad,\left \{ \begin{aligned} i=1, 2, \cdots, n \\ j=1, 2, \cdots, c \end{aligned} \right.$$

即：

$$\mu_{j}\left ( x_{i} \right) = \frac{\left [ 1/\left \| x_{i} - \mu_{j} \right \| ^{2} \right ]^{1/(b-1)}} {\displaystyle\sum_{k=1}^{c} \left [ 1/\left \| x_{i} - \mu_{k} \right \| ^{2} \right ]^{1/(b-1)}} \qquad\qquad,\left \{ \begin{aligned} i=1, 2, \cdots, n \\ j=1, 2, \cdots, c \end{aligned} \right.$$