1/0 -> linear regression can't model well -> -oo or +oo
probablity -> LR ok, but sigmoid function is better, S shape -> p = 1/1+e^-x = e^x/1+e^x
log(p/1-p) = bx -> logit function
Use maxmun likelihood function to solve
Loss on training data: L = sum(l(y, f(x))
1. Square loss = (y - f(x))^2
2. Logistic loss = Sum(yi ln(1 + e ^ -f(x)) + (1-yi) ln(1 + e ^ f(x)) + lamda* (w)^2
probablity -> LR ok, but sigmoid function is better, S shape -> p = 1/1+e^-x = e^x/1+e^x
log(p/1-p) = bx -> logit function
Use maxmun likelihood function to solve
Loss on training data: L = sum(l(y, f(x))
1. Square loss = (y - f(x))^2
2. Logistic loss = Sum(yi ln(1 + e ^ -f(x)) + (1-yi) ln(1 + e ^ f(x)) + lamda* (w)^2
No comments:
Post a Comment