Ridge regression is a regression method based on reducing the size of the regression coefficients, and thus increasing the bias of the model.

Like in OLS regresion, the cost function is an RSS function, but this time with an extra term pertaining to the shrinkage of each regression coefficient:

S(\beta,\lambda) = \text{RSS}(\beta) + \lambda \sum_{j=1}^p\beta_j^2 = \sum_{i=1}^N (y_i - \hat y_i)^2 + \lambda \sum_{j=1}^p\beta_j^2 ,

S(\beta, \lambda) = (\mathbf y - \mathbf X\beta)^T(\mathbf y - \mathbf X\beta) + \lambda \beta^T\beta

Here, \lambda \ge 0 is a parameter that controls how much the \beta's are shrunk'; the greater the \lambda, the more shrinkage. The Ridge estimator is of course now

\hat \beta^{\text{ridge}} = \underset{\beta}{\arg \min}\ S(\beta,\lambda) .

Like explained in the page about OLS, the above minimization problem can be written on matrix form as 1

\hat \beta^{\text{ridge}} = (\mathbf X^T\mathbf X + \lambda \mathbf I)^{-1}\mathbf X^T\mathbf y ,

where \mathbf I is the p\times p identity matrix. Notice that when adding a positive constant to the diagonal (or ridge 😉), the matrix inside the parentheses will always be invertible. This was one of the main motivating factors for Ridge regression. Its also worth noting that with orthonormal, inputs, Ridge regression is just a scaled version of the OLS regression, \hat \beta^{\text{ridge}} = \frac{1}{1 + \lambda}\hat \beta.

Now what is the whole idea behind shrinking the regression coefficients? Well, its quite clear that when $\lambda = 0$ the degrees of freedom of the fit is $\text{df}(\lambda) = p$, since this is the number of parameters. However, as $\lambda \rightarrow \infty$, then also $\text{df}(\lambda) \rightarrow 0$, since we gradually reduce the influence the different parameters have on the fit. So the goal is essentially to reduce the influence of unimportant parameters, and thus reduce [[Overfitting|overfitting]]. This ties directly in with reducing the [[Variance|variance]] by increasing the [[Bias|bias]], following the [[Bias-variance tradeoff]].

The Ridge regressor can be expressed in a closed-form expression, but for other regression methods, this might not be the case. Therefore, here is a more general formula: \hat \beta^{\text{ridge}} = \underset{\beta}{\arg \min} \begin{Bmatrix}\sum_{i=1}^N\left(y_i - \beta_0 - \sum_{j=1}^p x_{ij}\beta_j\right)^2 + \lambda \sum_{j=1}^p \beta_j^2 \end{Bmatrix} .

12:02 Wednesday 26 May 2021