Notes on Machine Learning 10: Bayesian linear regression
by 장승환
(ML 10) Baeysian linear regression
Linear regression is one of the most powerful general purpose tools.
Using MLE for ligression often leads to overfitting, which can be a sever problem.
Using MAP can solve the problem of overfitting, but we don’t have rpresentation of uncertainty in this case (not only in $w$ but in $y$).
Using Bayesian approach can optimize a appropriate loss function, and gives us the predictive distribution $p(y \vert x, D)$, which is what we really want.
Setup. Given data $D = ((x_1, y_1), \ldots, (x_n, y_n))$, $x_i \in \mathbb{R}^d$, $y_i \in \mathbb{R}$.
Model. $Y_1, \ldots, Y_n$ independent given $w$ with $Y_i \sim N(w^Tx_i, a^{-1})$, where $a = \frac{1}{\sigma^2} >0)$ is the precision and $w = (w_1, \ldots, w_d) \sim N(0, b^{-1}I)$ with $B >0$.
Assume $a, b$ are known and the parameter is given by $\theta = w$.
(ML 10.2) (ML 10.3) Posterior for linear regression
Remark. Can intoroduce nonlinearity (in $x$) by replacing $x_i$ by $\varphi(x_i) = (\varphi_1(x_i), \ldots, \varphi_m(x_i))$.
Likelihood.
(ML 10.4~7) Predictive distribution for linear regression
Subscribe via RSS