Modeling
Naive Bayes assumption: x 的所有 feature coordinates 是 conditionally independent 的
$P(x1,…,xM∣Ck)=P(x1∣Ck)⋯P(xM∣Ck)=∏j=1MP(xj∣Ck)$
When classifying, we can simply find the class Ck that maximizes P(Ck∣x) using the Bayes rule:
$argkmaxP(Ck∣x)=argkmaxP(Ck,x)=argkmaxP(Ck)P(x∣Ck)=argkmaxP(Ck)j=1∏MP(xj∣Ck)by Naive Bayes assumption$
For prior (class distribution): 仍然假设 class label 的 Probability 是 constant 的, 即 model 为 Bernoulli 分布
$y∼Bernulli(ϕ)即P(C_1)= \phi,P(C_0) = 1-\phi$