EM算法

通用EM

EM

• 通用EM
  • 目标函数：极大似然函数$logP(X|\theta )=log{\Sigma }_{z}P(x,z|\theta )$
  • 用于：不完整数据的对数似然函数
    • 不知Z的数据，只知道Z的后验分布$P(z|x,{\theta }^{old})$
    • 考虑其期望$Q(\theta ,{\theta }^{old})=E_{p(z|x,{\theta }^{old})}(logP(x,z|\theta ))$
    • 最大化期望${\theta }^{new}=argma{x}_{\theta }Q(\theta ,{\theta }^{old})$
  • E：求$P(z|x,{\theta }^{old})$
  • M：${\theta }^{new}=argma{x}_{\theta }Q(\theta ,{\theta }^{old})$
    • why是启发式的，但却存在似然函数？
      • $Q(\theta ,{\theta }^{old})=E_{p(z|x,{\theta }^{old})}(logP(x,z|\theta ))=p(x;\theta )$
  • 完整数据和不完整数据的比较
  • 不完整数据：$logp(x)={\Sigma }_{i}log{\Sigma }_{z}p(x_i|z)p(z)={\Sigma }_{i}log{\Sigma }_{k=1}^K{\pi }_{k}N(x_i|{\mu }_k,{\Sigma }_k)$
    • 不完整数据中，参数之间是耦合的，不存在封闭解
  • 完整数据
    • $logp(x,z|\theta )=logp(z|\theta )p(x|z,\theta )={\Sigma }_i{\Sigma }_k{z}_{ik}(log{\pi }_k+logN(x_i|{\mu }_k,{\Sigma }_k))$
    • $$\begin{gathered} E_z(logp(x,z|\theta )) \\ ={\Sigma }_i{\Sigma }_{k}E(z_{ik})(log{\pi }_k+logN(x_i|{\mu }_k,{\Sigma }_k)) \\ ={\Sigma }_i{\Sigma }_k\gamma (z_{ik})(log{\pi }_k+logN(x_i|{\mu }_k,{\Sigma }_k)) \end{gathered}$$

	E	M	目标函数	对谁求导
通用(z)	$P(z\Vert x,{\theta }^{old})$	$\theta^{new}=argmax_\theta Q(\theta,\theta^{old})=log(p(x	\theta))$	$Q(\theta ,{\theta }^{old})=E_{p(z\Vert x,{\theta }^{old})}(logP(x,z\Vert \theta ))$
GMM(y)	$$\begin{gathered} \gamma (z_{ik}) \\ =p(z_{ik=1}\Vert x_i) \\ =\frac{p(z_{ik}=1)p(x_i\Vert z_k=1)}{{\Sigma }_{k=1}^{K}p(z_{ik}=1)p(x_i\Vert z_k=1)} \\ =\frac{{\pi }_{k}N(x_i\Vert {\mu }_k,{\Sigma }_k)}{{\Sigma }_{k=1}^K{\pi }_{k}N(x_i\Vert {\mu }_k,{\Sigma }_k)} \end{gathered}$$	$$\begin{gathered} {\mu }_k=\frac{{\Sigma }_i\gamma (z_{ik})x_i}{{\Sigma }_i\gamma (z_{ik})} \\ {\pi }_k=\frac{{\Sigma }_i\gamma (z_{ik})}{N} \\ {\Sigma }_k=\frac{{\Sigma }_i\gamma (z_{ik})(x_i-{\mu }_k)(x_i-{\mu }_k{)}^T}{\gamma (z_{ik})} \end{gathered}$$	$p(x;\theta )={\Pi }_i^N{\Sigma }_{k=1}^K{\pi }_{k}N(x_i\Vert {\mu }_k,{\Sigma }_k),其中{\Sigma }_k{\pi }_k=1,0\le {\pi }_k\le 1,Q(\theta ,{\theta }^{old})=log({\Sigma }_{z}P(x,z\Vert \theta ))=logp(x\Vert \theta )$	$logp(x\Vert \theta )$
HMM (y)	$$\begin{gathered} \xi (y_t,y_{t+1})=P(y_t,y_{t+1}\Vert x)=\frac{\alpha (y_t)P(x_{t+1}\Vert y_{t+1})\beta (y_{t+1})a_{y_{t+1},y_t}}{p(x)} \\ {\gamma }_t^i=P(y_t^i=1\Vert x,{\theta }^p) \\ E(n_{ij}\Vert x,{\theta }^p)={\Sigma }_{t=1}^T{\gamma }_t^i{x}_t^i \\ E(m_{ij}\Vert x,{\theta }^p)={\Sigma }_{t=1}^{T-1}{\xi }_{t,t+1}^{ij} \end{gathered}$$	$$\begin{gathered} {\hat{\alpha }}_{ij}=\frac{m_{ij}}{{\Sigma }_{k=1}^N{m}_{ik}} \\ {\hat{\eta }}_{ij}=\frac{n_{ij}}{{\Sigma }_{k=1}^N{n}_{ik}} \\ {\hat{\pi }}_i=y_1^i \end{gathered}$$	$$\begin{gathered} Q(\theta ,{\theta }^{old})=E_{p(y\Vert x,{\theta }^{old})}(logP(x,y\Vert \theta ))={\Sigma }_{y}p(y\Vert x,{\theta }^{old})logP(x,y\Vert \theta ) \\ ={\Sigma }_y((log{\pi }_i+{\Sigma }_{t=1}^{T-1}log{a}_{t,t+1}+{\Sigma }_{t=1}^{T}P(x_t\Vert y_t))P(x,y\Vert {\theta }^{old})) \\ s.t{\Sigma }_{i=1}^N{\pi }_i=1 \end{gathered}$$	$L=Q(\theta ,{\theta }^{old})+\lambda ({\Sigma }_{i=1}^N{\pi }_i-1)$

EM收敛性保证

• 目标：最大化$P(x|\theta )={\Sigma }_{z}p(x,z|\theta )$
  • 直接优化$P(x|\theta )$很困难，但优化完整数据的$p(x,z|\theta )$容易
• 证明
  • 分解
  • 对任意分布q(z),下列分解成立
    • $$\begin{gathered} lnp(x|\theta )=L(q,\theta )+KL(q||p) \\ 其中， \\ L(q,\theta )={\Sigma }_{z}q(z)ln(\frac{p(x,z|\theta )}{q(z)}) \\ KL(q||p)=-{\Sigma }_{z}q(z)ln(\frac{p(z|x,\theta )}{q(z)}) \\ KL(q||p)\ge 0,L(q,\theta )是lnp(x|\theta )的下界 \end{gathered}$$
  • E：$$\begin{gathered} 最大化L(q,\theta ), \\ q(z)=P(z|x,{\theta }^{old}) \end{gathered}$$
  • $M:原来的下界L(q,\theta )={\Sigma }_{z}P(z|x,{\theta }^{old})ln(\frac{p(x,z|\theta )}{q(z)})=Q(\theta ,{\theta }^{old})+const---正好是期望$
  • 下界提升了
    
    
    
    https://www.bilibili.com/video/av31906558?from=search&seid=2112421761429235163

GMM(聚类

5.2GMM高斯混合模型和EM

• 概率解释: 假设有K个簇，每一个簇服从高斯分布，以概率π𝑘随机选择一个簇 k ，从其分布中采样出一个样本点，如此得到观测数据
• N个样本点𝒙的似然函数(Likelihood)
  • $p(x;\theta )={\Pi }_i^N{\Sigma }_{k=1}^K{\pi }_{k}N(x_i|{\mu }_k,{\Sigma }_k),其中{\Sigma }_k{\pi }_k=1,0\le {\pi }_k\le 1$
  • 引入隐变量，指示所属类,k维独热表示
    • $p(z_k=1)={\pi }_k$
    • $p(x_i|z)={\Pi }_k^{K}N(x_i|{\mu }_k,{\Sigma }_k{)}^{z_k}$
      • $p(x_i|z_k=1)=N(x_i|{\mu }_k,{\Sigma }_k)$
    • $p(x_i)={\Sigma }_{z}p(x_i|z)p(z)={\Sigma }_{k=1}^K{\pi }_{k}N(x_i|{\mu }_k,{\Sigma }_k)$
• 从属度(可以看做，xi属于第k个簇的解释
  • $$\begin{gathered} \gamma (z_{ik}) \\ =p(z_{ik=1}|x_i) \\ =\frac{p(z_{ik}=1)p(x_i|z_k=1)}{{\Sigma }_{k=1}^{K}p(z_{ik}=1)p(x_i|z_k=1)} \\ =\frac{{\pi }_{k}N(x_i|{\mu }_k,{\Sigma }_k)}{{\Sigma }_{k=1}^K{\pi }_{k}N(x_i|{\mu }_k,{\Sigma }_k)} \end{gathered}$$

参数学习：极大似然估计–EM

• 极大似然估计
  • 难：log里面有求和，所有参数耦合
  • 似然函数取最大值时满足的条件：$log(P(x|\theta ))=log({\Sigma }_{z}P(x,z|\theta )对{\mu }_k求导$
    • $0=-{\Sigma }_{i=1}^N\frac{{\pi }_{k}N(x_i|{\mu }_k,{\Sigma }_k)}{{\Sigma }_{k=1}^K{\pi }_{k}N(x_i|{\mu }_k,{\Sigma }_k)}{\Sigma }_k^{-1}(x_i-{\mu }_k)$
      • ${\mu }_k=\frac{{\Sigma }_i\gamma (z_{ik})x_i}{\gamma (z_{ik})}$
      • ${\pi }_k=\frac{{\Sigma }_i\gamma (z_{ik})}{N}$
      • ${\Sigma }_k=\frac{{\Sigma }_i\gamma (z_{ik})(x_i-{\mu }_k)(x_i-{\mu }_k{)}^T}{\gamma (z_{ik})}$
    • 这不是封闭解–》EM
      • E：给定当前参数估计值，求后验概率$\gamma (z_{ik})=E(z_{ik})$
      • M：依据后验概率$\gamma (z_{ik})$，求参数估计${\mu }_k、{\pi }_k、{\Sigma }_k$
      • 迭代收敛到局部极小

HMM

4.1.3 学习，参数估计

• 极大似然估计：EM算法

  • 最大化$P(x|\theta )$
  • 参数$A、\pi ,输出分布的参数$

• $P(x|\theta )={\Sigma }_{y_1,y_2,...,y_T}P(x,y)={\Sigma }_{y_1,y_2,...,y_T}{\pi }_{y_1}{\Pi }_{t=1}^{T-1}{a}_{y_{t+1},y_t}{\Pi }_{t=1}^{T}P(x_t|y_t,\eta )$

• $假设P(x_t|y_t,\eta )={\Pi }_{i=1}^M{\Pi }_{j=1}^L[{\eta }_{ij}{]}^{y_t^i{x}_t^j}$

• M
  

• $$\begin{gathered} {\hat{\alpha }}_{ij}=\frac{m_{ij}}{{\Sigma }_{k=1}^N{m}_{ik}} \\ {\hat{\eta }}_{ij}=\frac{n_{ij}}{{\Sigma }_{k=1}^N{n}_{ik}} \\ {\hat{\pi }}_i=y_1^i \end{gathered}$$

• E步

• $$\begin{gathered} \xi (y_t,y_{t+1})=P(y_t,y_{t+1}|x) \\ =\frac{P(x|y_t,y_{t+1})P(y_{t+1}|y_t)P(y_t)}{p(x)} \\ =\frac{P(x1,...x_t|y_t)P(x_{t+1}|y_{t+1})P(x_{t+2},...x_n|y_{t+1})P(y_{t+1}|y_t)P(y_t)}{p(x)} \\ =\frac{\alpha (y_t)P(x_{t+1}|y_{t+1})\beta (y_{t+1})a_{y_{t+1},y_t}}{p(x)} \end{gathered}$$

• ${\gamma }_t^i=P(y_t^i=1|x,{\theta }^p)$

• $$\begin{gathered} E(n_{ij}|x,{\theta }^p)={\Sigma }_{t=1}^T{\gamma }_t^i{x}_t^i \\ E(m_{ij}|x,{\theta }^p)={\Sigma }_{t=1}^{T-1}{\xi }_{t,t+1}^{ij} \end{gathered}$$
  
  

• $$\begin{gathered} Q(\theta ,{\theta }^{old})=E_{p(y|x,{\theta }^{old})}(logP(x,y|\theta ))={\Sigma }_{I}p(y|x,{\theta }^{old}logP(x,y|\theta ) \\ ={\Sigma }_I((log{\pi }_i+{\Sigma }_{t=1}^{T-1}log{a}_{t,t+1}+{\Sigma }_{t=1}^{T}P(x_t|y_t))P(x,y|{\theta }^{old})) \\ s.t{\Sigma }_{i=1}^N{\pi }_i=1 \end{gathered}$$

• 缺点

  • 仅捕捉了状态之间和状态及其对应输出之间的关系（上下文）
  • 学习目标和预测目标不匹配
    • 我们只要p(y|x),但只知道p(x,y)—产生式模型

PLSA

LDA