Generalized linear model

why GLM

  1. extend linear model to multi-classification
  2. generize the maximum likehood procedure to a generic distribution (exponential family distribution)

exponential family distribution

$p(y_n|\eta) := exp[y_n \eta_n - A(\eta_n)] h(y)$
bernoulli distribution, Gaussian distribution and multinomial distribution can be derived into the form.

several things keep in mind about logistic regression

how logistic regression come? what assumption is made?

  • binary output: y equals to 1 or 0
  • assume $y_n$ is independent of each others
  • probabilistic model is applied: $P(y=1|\mathbf{x},\mathbf{beta})$ and $P(y=0|\mathbf{x},\mathbf{beta})$ they are in fact:
    $$P(y{predict}=y{true}|\mathbf{x},\mathbf{beta})$$
  • log-likehood is $$log(P(y{predict}=y{true}|\mathbf{x},\mathbf{beta})) = log(fun(\mathbf{\beta})) = L(\mathbf{\beta})$$
  • pay attention to diffence of $p(y_n|\eta_n)$ and $p(\mathbf{y}|\mathbf{\eta})$
  • it is indeed bernoulli distribution