1
$\begingroup$

Consider the following ridge regression problem: minimize the loss function $\sum_{i=1}^n ||y_i - w^T x_i||_2^2 + \lambda ||w||_2^2$ with respect to the weight vector w. Taking derivative with respect to w, I get $\sum_{i=1}^n 2(y_i - w^T x_i)(-x_i) + 2\lambda w$ which implies $w =(\sum_{i=1}^n (y_i - w^T x_i)(x_i)) / 2\lambda $. Is this wrong? I know that the solution is $(X^TX - \lambda I)^{-1}X^Ty$.

$\endgroup$
3
$\begingroup$

Your derivative is okay. Just remember to put all the $w$-terms on the same side of the equation $$\eqalign{ \sum_i x_i y_i &= \lambda w + \sum_i x_i x_i^Tw \cr }$$ Then pull $w$ out of the summation, since it's independent of $i$ $$\eqalign{ \sum_i y_i x_i &= \Big(\lambda I + \sum_i x_ix_i^T\Big)w \cr }$$ At this point, dispose of the summations in favor of matrix notation $$\eqalign{ X^Ty &= \big(\lambda I + X^TX\big)w \cr }$$ where $x_i$ is the $i^{th}$ column of $X,\,$ and $\,y_i$ is the $i^{th}$ component of $y$.

$\endgroup$
1
$\begingroup$

Your solution has $w$ on both sides of the equation, and furthermore, $w$ is inside the summation, which is a problem.

I recommend taking a look at section 2 of Andrew Ng's CS 229 course notes. In short, using matrix math allows you to use properties of the gradient of a trace, which allows for a straightforward derivation. Note that his course notes are for standard least squares, but getting ridge regression is basically the same.

$\endgroup$
  • 1
    $\begingroup$ 1. Please don't have your answer rely on a link which may disappear at any time. Describe - or even quote - the essential parts (giving proper credit and including a complete reference if possible), so that when that link no longer works your answer doesn't become invalidated. 2. While I believe I understand exactly the sense in which your last sentence is intended, I doubt the OP nor many of our readers would see the immediate connection. Perhaps you could the details there (or perhaps link to some of the other answers on site which would be sufficient to give the right sense). $\endgroup$ – Glen_b Jan 21 '16 at 6:00
  • $\begingroup$ Thanks for the helpful comment - much better than just downvoting me. At this point others have contributed good answers - I'll do it differently next time. $\endgroup$ – Mageek Jan 21 '16 at 17:57
1
$\begingroup$

I think you are asking something different to the question title than in your post body.

Regarding the question in the title:

Most algorithms needed to be coded. Using matrix notation greatly simplifies the coding as well as emphasizes the points that the algorithm can be immediately vectorized (take advantage of linear algebra libraries). Also less indexing usually translates in less typing which is always welcome. :)

Regarding the question in the post:

  1. Usually one expresses this cost function with a $\frac{1}{2}$ scalar ahead of it exactly to get rid of the $2$'s in the expression.
  2. Usually the expression is $(X^T X + \lambda I)^{-1}X^Ty$ (notice the sign of $\lambda$). You use "$+$" in your original expression but "$-$" afterwards.
  3. See the great answers by whuber and Brian Borchers here and here respectively. They use matrix notation to derive the ridge regression problem.
  4. You essentially want to take advantage of the following notational property to go from scalar to matrix notation: $\sum_{i}^n (y_i - X_i w)^2 = (y -Xw)^T (y-Xw)$. (Similarly $\lambda ||w||_2^2 = \lambda w^T w$.)
  5. No you are not wrong. You are really close; just move a sign out of the summation and you are almost there! You want your derivative to equate $-2X^T(y-Xw) + 2\lambda w$ (in matrix notation).
$\endgroup$
  • 1
    $\begingroup$ I used $w$ to match your notation; please use $\beta$ in general it is much more canonical. $\endgroup$ – usεr11852 Jan 21 '16 at 6:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.