逻辑回归模型中，f(x)由线性函数定义，包括参数β和特征x。该模型计算出的P(y=1)是事件发生的概率，其可解释性在于当特征值增加1时，对概率的影响。通过对数似然函数，可以理解模型如何将输入转换为预测概率，并且展示了特征值变化如何影响f(x)的值。
Interpretability of Logistic regression
Rewrite the function
-
original function:
l o g i s t i c ( f ( x ) ) = 1 1 + e x p ( − f ( x ) ) logistic(f(x)) = \frac{1}{1+exp(-f(x))} logistic(f(x))=1+exp(−f(x))1 -
By definition
f ( x ) = β 0 + β 1 x 1 + . . . β p x p f(x) =\beta_0 +\beta_{1}x_1+...\beta_{p}x_{p} f(x)=β0+β1x1+...βpxp
Thus,
P ( y = 1 ) = 1 1 + e x p ( − ( β 0 + β 1 x 1 + . . . β p x p ) ) P(y=1)=\frac{1}{1+exp(-(\beta_0 +\beta_{1}x_1+...\beta_{p}x_{p}))} P(y=1)=1+exp(−(β0+β1x1+...βpxp))1 -
l n ( P ( y = 1 ) 1 − P ( y = 1 ) ) = l o g ( P ( y = 1 ) P ( y = 0 ) ) ln(\frac{P(y=1)}{1-P(y=1)}) = log(\frac{P(y=1)}{P(y=0)}) ln(1−P(y=1)P(y=1))=log(P(y=0)P(y=1))
= l o g ( 1 1 + e x p ( − f ( x ) ) / e x p ( − f ( x ) ) 1 + e x p ( − f ( x ) ) ) = l o g ( 1 e x p ( − f ( x ) ) ) = l o g 1 − l o g ( e x p ( − f ( x ) ) ) = f ( x ) log(\frac{1}{1+exp(-f(x))}/\frac{exp(-f(x))}{1+exp(-f(x))}) = log(\frac{1}{exp(-f(x))}) = log1 - log(exp(-f(x))) =f(x) log(1+exp(−f(x))1/1+exp(−f(x))exp(−f(x)))=log(exp(−f(x))1)=log1−log(exp(−f(x)))=f(x)
Interpretability
increase the feature value by 1
387
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
