1. Introduction
为了解决高斯模型的单峰性的问题,我们引入多个高斯模型的加权平均来拟合多峰数据:
p
(
x
)
=
∑
k
=
1
K
α
k
N
(
μ
k
,
Σ
k
)
p(x)=\sum\limits_{k=1}^K\alpha_k\mathcal{N}(\mu_k,\Sigma_k)
p(x)=k=1∑KαkN(μk,Σk) 引入隐变量
z
z
z,这个变量表示对应的样本
x
x
x 属于哪一个高斯分布,这个变量是一个离散的随机变量:
p
(
z
=
i
)
=
p
i
∑
i
=
1
k
p
(
z
=
i
)
=
1
p(z=i)=p_i\\ \sum\limits_{i=1}^kp(z=i)=1
p(z=i)=pii=1∑kp(z=i)=1 作为一个生成式模型,高斯混合模型通过隐变量
z
z
z 的分布来生成样本。用概率图来表示:
其中,节点
z
z
z 就是上面的概率,
x
x
x 就是生成的高斯分布。于是对
p
(
x
)
p(x)
p(x):
p
(
x
)
=
∑
z
p
(
x
,
z
)
=
∑
k
=
1
K
p
(
x
,
z
=
k
)
=
∑
k
=
1
K
p
(
z
=
k
)
p
(
x
∣
z
=
k
)
p(x)=\sum\limits_zp(x,z)=\sum\limits_{k=1}^Kp(x,z=k)=\sum\limits_{k=1}^Kp(z=k)p(x|z=k)
p(x)=z∑p(x,z)=k=1∑Kp(x,z=k)=k=1∑Kp(z=k)p(x∣z=k) 因此:
p
(
x
)
=
∑
k
=
1
K
p
k
N
(
x
∣
μ
k
,
Σ
k
)
p(x)=\sum\limits_{k=1}^Kp_k\mathcal{N}(x|\mu_k,\Sigma_k)
p(x)=k=1∑KpkN(x∣μk,Σk)
2.极大似然估计
样本为 X = ( x 1 , x 2 , ⋯ , x N ) X=(x_1,x_2,\cdots,x_N) X=(x1,x2,⋯,xN), ( X , Z ) (X,Z) (X,Z) 为完全参数,参数为 θ = p 1 , p 2 , ⋯ , p K , μ 1 , μ 2 , ⋯ , μ K Σ 1 , Σ 2 , ⋯ , Σ K \theta={p_1,p_2,\cdots,p_K,\mu_1,\mu_2,\cdots,\mu_K\Sigma_1,\Sigma_2,\cdots,\Sigma_K} θ=p1,p2,⋯,pK,μ1,μ2,⋯,μKΣ1,Σ2,⋯,ΣK。我们通过极大似然估计得到 θ \theta θ 的值: θ M L E = a r g m a x θ log p ( X ) = a r g m a x θ ∑ i = 1 N log p ( x i ) = a r g m a x θ ∑ i = 1 N log ∑ k = 1 K p k N ( x i ∣ μ k , Σ k ) \begin{aligned}\theta_{MLE}&=\mathop{argmax}\limits_{\theta}\log p(X)=\mathop{argmax}{\theta}\sum_{i=1}^N\log p(x_i)\\ &=\mathop{argmax}\limits_{\theta}\sum_{i=1}^N\log \sum_{k=1}^Kp_k\mathcal{N}(x_i|\mu_k,\Sigma_k) \end{aligned} θMLE=θargmaxlogp(X)=argmaxθi=1∑Nlogp(xi)=θargmaxi=1∑Nlogk=1∑KpkN(xi∣μk,Σk) 这个表达式直接通过求导,由于连加号的存在,无法得到解析解。因此需要使用 EM 算法。
3. EM求解GMM
EM 算法的基本表达式为: θ t + 1 = a r g m a x θ E z ∣ x , θ t [ p ( x , z ∣ θ ) ] \theta^{t+1}=\mathop{argmax}\limits_{\theta}\mathbb{E}_{z|x,\theta_t}[p(x,z|\theta)] θt+1=θargmaxEz∣x,θt[p(x,z∣θ)]。套用 GMM 的表达式,对数据集来说: Q ( θ , θ t ) = ∑ z [ log ∏ i = 1 N p ( x i , z i ∣ θ ) ] ∏ i = 1 N p ( z i ∣ x i , θ t ) = ∑ z [ ∑ i = 1 N log p ( x i , z i ∣ θ ) ] ∏ i = 1 N p ( z i ∣ x i , θ t ) \begin{aligned}Q(\theta,\theta^t)&=\sum\limits_z[\log\prod\limits_{i=1}^Np(x_i,z_i|\theta)]\prod \limits_{i=1}^Np(z_i|x_i,\theta^t)\\ &=\sum\limits_z[\sum\limits_{i=1}^N\log p(x_i,z_i|\theta)]\prod \limits_{i=1}^Np(z_i|x_i,\theta^t) \end{aligned} Q(θ,θt)=z∑[logi=1∏Np(xi,zi∣θ)]i=1∏Np(zi∣xi,θt)=z∑[i=1∑Nlogp(xi,zi∣θ)]i=1∏Np(zi∣xi,θt) 对于中间的那个求和号,展开,第一项为: ∑ z log p ( x 1 , z 1 ∣ θ ) ∏ i = 1 N p ( z i ∣ x i , θ t ) = ∑ z log p ( x 1 , z 1 ∣ θ ) p ( z 1 ∣ x 1 , θ t ) ∏ i = 2 N p ( z i ∣ x i , θ t ) = ∑ z 1 log p ( x 1 , z 1 ∣ θ ) p ( z 1 ∣ x 1 , θ t ) ∑ z 2 , ⋯ , z K ∏ i = 2 N p ( z i ∣ x i , θ t ) = ∑ z 1 log p ( x 1 , z 1 ∣ θ ) p ( z 1 ∣ x 1 , θ t ) \begin{aligned} \sum\limits_z\log p(x_1,z_1|\theta)\prod\limits_{i=1}^Np(z_i|x_i,\theta^t)&=\sum\limits_z\log p(x_1,z_1|\theta)p(z_1|x_1,\theta^t)\prod\limits_{i=2}^Np(z_i|x_i,\theta^t)\\ &=\sum\limits_{z_1}\log p(x_1,z_1|\theta) p(z_1|x_1,\theta^t)\sum\limits_{z_2,\cdots,z_K}\prod\limits_{i=2}^Np(z_i|x_i,\theta^t)\\ &=\sum\limits_{z_1}\log p(x_1,z_1|\theta)p(z_1|x_1,\theta^t)\end{aligned} z∑logp(x1,z1∣θ)i=1∏Np(zi∣xi,θt)=z∑logp(x1,z1∣θ)p(z1∣x1,θt)i=2∏Np(zi∣xi,θt)=z1∑logp(x1,z1∣θ)p(z1∣x1,θt)z2,⋯,zK∑i=2∏Np(zi∣xi,θt)=z1∑logp(x1,z1∣θ)p(z1∣x1,θt) 类似地, Q Q Q 可以写为: Q ( θ , θ t ) = ∑ i = 1 N ∑ z i log p ( x i , z i ∣ θ ) p ( z i ∣ x i , θ t ) Q(\theta,\theta^t)=\sum\limits_{i=1}^N\sum\limits_{z_i}\log p(x_i,z_i|\theta)p(z_i|x_i,\theta^t) Q(θ,θt)=i=1∑Nzi∑logp(xi,zi∣θ)p(zi∣xi,θt) 对于 p ( x , z ∣ θ ) p(x,z|\theta) p(x,z∣θ): p ( x , z ∣ θ ) = p ( z ∣ θ ) p ( x ∣ z , θ ) = p z N ( x ∣ μ z , Σ z ) p(x,z|\theta)=p(z|\theta)p(x|z,\theta)=p_z\mathcal{N}(x|\mu_z,\Sigma_z) p(x,z∣θ)=p(z∣θ)p(x∣z,θ)=pzN(x∣μz,Σz) 对 p ( z ∣ x , θ t ) p(z|x,\theta^t) p(z∣x,θt): p ( z ∣ x , θ t ) = p ( x , z ∣ θ t ) p ( x ∣ θ t ) = p z t N ( x ∣ μ z t , Σ z t ) ∑ k p k t N ( x ∣ μ k t , Σ k t ) p(z|x,\theta^t)=\frac{p(x,z|\theta^t)}{p(x|\theta^t)}=\frac{p_z^t\mathcal{N}(x|\mu_z^t,\Sigma_z^t)}{\sum\limits_kp_k^t\mathcal{N}(x|\mu_k^t,\Sigma_k^t)} p(z∣x,θt)=p(x∣θt)p(x,z∣θt)=k∑pktN(x∣μkt,Σkt)pztN(x∣μzt,Σzt) 代入 Q Q Q: Q = ∑ i = 1 N ∑ z i log p z i N ( x i ∣ μ z i , Σ z i ) p z i t N ( x i ∣ μ z i t , Σ z i t ) ∑ k p k t N ( x i ∣ μ k t , Σ k t ) Q=\sum\limits_{i=1}^N\sum\limits_{z_i}\log p_{z_i}\mathcal{N(x_i|\mu_{z_i},\Sigma_{z_i})}\frac{p_{z_i}^t\mathcal{N}(x_i|\mu_{z_i}^t,\Sigma_{z_i}^t)}{\sum\limits_kp_k^t\mathcal{N}(x_i|\mu_k^t,\Sigma_k^t)} Q=i=1∑Nzi∑logpziN(xi∣μzi,Σzi)k∑pktN(xi∣μkt,Σkt)pzitN(xi∣μzit,Σzit) 下面需要对 Q Q Q 值求最大值: Q = ∑ k = 1 K ∑ i = 1 N [ log p k + log N ( x i ∣ μ k , Σ k ) ] p ( z i = k ∣ x i , θ t ) Q=\sum\limits_{k=1}^K\sum\limits_{i=1}^N[\log p_k+\log \mathcal{N}(x_i|\mu_k,\Sigma_k)]p(z_i=k|x_i,\theta^t) Q=k=1∑Ki=1∑N[logpk+logN(xi∣μk,Σk)]p(zi=k∣xi,θt)
-
p
k
t
+
1
p_k^{t+1}
pkt+1:
p k t + 1 = a r g m a x p k ∑ k = 1 K ∑ i = 1 N [ log p k + log N ( x i ∣ μ k , Σ k ) ] p ( z i = k ∣ x i , θ t ) s . t . ∑ k = 1 K p k = 1 p_k^{t+1}=\mathop{argmax}\limits_{p_k}\sum\limits_{k=1}^K\sum\limits_{i=1}^N[\log p_k+\log \mathcal{N}(x_i|\mu_k,\Sigma_k)]p(z_i=k|x_i,\theta^t)\\ s.t.\ \sum\limits_{k=1}^Kp_k=1 pkt+1=pkargmaxk=1∑Ki=1∑N[logpk+logN(xi∣μk,Σk)]p(zi=k∣xi,θt)s.t. k=1∑Kpk=1 即: p k t + 1 = a r g m a x p k ∑ k = 1 K ∑ i = 1 N log p k p ( z i = k ∣ x i , θ t ) s . t . ∑ k = 1 K p k = 1 p_k^{t+1}=\mathop{argmax}\limits_{p_k}\sum\limits_{k=1}^K\sum\limits_{i=1}^N\log p_kp(z_i=k|x_i,\theta^t)\\ s.t.\ \sum\limits_{k=1}^Kp_k=1 pkt+1=pkargmaxk=1∑Ki=1∑Nlogpkp(zi=k∣xi,θt)s.t. k=1∑Kpk=1 引入 Lagrange 乘子: L ( p k , λ ) = ∑ k = 1 K ∑ i = 1 N log p k p ( z i = k ∣ x i , θ t ) − λ ( 1 − ∑ k = 1 K p k ) L(p_k,\lambda)=\sum\limits_{k=1}^K\sum\limits_{i=1}^N\log p_kp(z_i=k|x_i,\theta^t)-\lambda(1-\sum\limits_{k=1}^Kp_k) L(pk,λ)=k=1∑Ki=1∑Nlogpkp(zi=k∣xi,θt)−λ(1−k=1∑Kpk)。所以: ∂ ∂ p k L = ∑ i = 1 N 1 p k p ( z i = k ∣ x i , θ t ) + λ = 0 ⇒ ∑ k ∑ i = 1 N 1 p k p ( z i = k ∣ x i , θ t ) + λ ∑ k p k = 0 ⇒ λ = − N \frac{\partial}{\partial p_k}L=\sum\limits_{i=1}^N\frac{1}{p_k}p(z_i=k|x_i,\theta^t)+\lambda=0\ \Rightarrow \sum\limits_k\sum\limits_{i=1}^N\frac{1}{p_k}p(z_i=k|x_i,\theta^t)+\lambda\sum\limits_kp_k=0\\ \Rightarrow\lambda=-N ∂pk∂L=i=1∑Npk1p(zi=k∣xi,θt)+λ=0 ⇒k∑i=1∑Npk1p(zi=k∣xi,θt)+λk∑pk=0⇒λ=−N
于是有: p k t + 1 = 1 N ∑ i = 1 N p ( z i = k ∣ x i , θ t ) p_k^{t+1}=\frac{1}{N}\sum\limits_{i=1}^Np(z_i=k|x_i,\theta^t) pkt+1=N1i=1∑Np(zi=k∣xi,θt)
- μ k , Σ k \mu_k,\Sigma_k μk,Σk,这两个参数是无约束的,直接求导即可。