本节记录一下由贝叶斯定理延伸出来的几种预测性建模的方法,主要为线性判别分析(一次,二次),朴素贝叶斯(稍稍提一下贝叶斯网络)
1. 判别分析
判别分析适用于自变量连续,因变量为分类型的情形;
- 设因变量 Y Y Y一共有 K K K个类别; ϵ l = P ( Y = l ) \epsilon_l=P(Y=l) ϵl=P(Y=l)表示类别 l l l的先验概率,满足 ∑ l = 1 K ϵ l = 1 \sum^K_{l=1}\epsilon_l=1 ∑l=1Kϵl=1; f l ( x ) = f ( x ∣ Y = l ) f_l(\bold{x})=f(\bold{x}|Y=l) fl(x)=f(x∣Y=l)表示类别 Y = l Y=l Y=l的观测下自变量 X = ( X 1 , ⋯ , X p ) \bold{X}=(X_1,\cdots,X_p) X=(X1,⋯,Xp)的概率密度函数;
由贝叶斯公式
P ( Y = l ∣ X = x ) = P ( Y = l ) f ( x ∣ Y = l ) ∑ i = 1 K P ( Y = i ) f ( x ∣ Y = i ) = ϵ l f l ( x ) ∑ i = 1 K ϵ i f i ( x ) P(Y=l|\bold{X}=\bold{x})=\frac{P(Y=l)f(\bold{x}|Y=l)}{\sum^K_{i=1}P(Y=i)f(\bold{x}|Y=i)}=\frac{\epsilon_lf_l(\bold{x})}{\sum^K_{i=1}\epsilon_if_i(\bold{x})} P(Y=l∣X=x)=∑i=1KP(Y=i)f(x∣Y=i)P(Y=l)f(x∣Y=l)=∑i=1Kϵifi(x)ϵlfl(x)
则预测 x \bold{x} x的类别 l ∗ l^* l∗为
l ∗ = a r g m a x l P ( Y = l ∣ X = x ) = a r g m a x l ϵ l f l ( x ) l^* = \underset{l}{argmax}\ P(Y=l|\bold{X}=\bold{x}) = \underset{l}{argmax}\ \epsilon_lf_l(\bold{x}) l∗=largmax P(Y=l∣X=x)=largmax ϵlfl(x)
- 线性判别分析的假设为,观测自变量满足多元正态分布,即 f l ( x ) ∼ M V N ( μ l , Σ l ) f_l(\bold{x})\sim MVN(\bold{\mu}_l,\bold{\Sigma}_l) fl(x)∼MVN(μl,Σl)
1.1 线性判别分析
- 在上述假设中进一步假设所有类别的协方差矩阵都相等: Σ l = Σ , l = 1 , ⋯ , K \Sigma_l=\Sigma,l=1,\cdots,K Σl=Σ,l=1,⋯,K
前述已经将预测类别的类转化为求 a r g m a x l ϵ l f l ( x ) \underset{l}{argmax}\ \epsilon_lf_l(\bold{x}) largmax ϵlfl(x),我们稍微计算一下
l o g ( ϵ l f l ( x ) ) = l o g [ ϵ l 1 ( 2 π ) p 2 ∣ Σ ∣ 1 2 e x p { − 1 2 ( x − μ l ) T Σ − 1 ( x − μ l ) } ] = δ l ( x ) + A \begin{aligned} log(\epsilon_lf_l(\bold{x}))&=log[\epsilon_l\frac{1}{(2\pi)^{\frac{p}{2}}|\Sigma|^{\frac{1}{2}}}exp\{-\frac{1}{2}(\bold{x}-\bold{\mu}_l)^T\Sigma^{-1}(\bold{x}-\bold{\mu}_l)\}]\\ &= \delta_l(\bold{x})+A\\ \end{aligned} log(ϵlfl(x))=log[ϵl(2π)2p∣Σ∣211exp{
−21(x−μl)TΣ−1(x−μl)}]=δl(x)+A