0
$\begingroup$

I am struggling with this point (since my Math is abysmal). I have looked at various YouTube videos (most of them just gloss over this or the Math is beyond me) and read various blog posts (same problem).

Basically, all I want to understand is: Why do the equations for the hyperplanes that the support vectors are on look like this?

$\vec{w}^T\vec{x}+b=1$

$\vec{w}^T\vec{x}+b=-1$

Why 1/-1?

I understand that for points on the hyperplane, this is true:

$\vec{w}^T\vec{x}+b=0$

For points "above" or in the direction of $\vec{w}$ it's $ > 0 $, for the other case it's $ < 0 $

I just don't really understand where the 1/-1 come from. Form what I have read, it has to do with scaling and ultimately choosing 1 is arbitrary, but I don't get it.

What is the problem with the scaling and I kind of makes the optimization problem easier, but the penny just won't drop.

I understand that by doing $c\vec{w}^T\vec{x}+cb=0$, I get the same hyperplane, but I don't understand why because of that I can just go: $\vec{w}^T\vec{x}+b=1$.

So in my head I am picturing this:

The equation for the support vectors are $\vec{w}^T\vec{x}+b= +d / -d $. I basically make two copies of my hyperplane and move them upwards/downwards until each copy hits the first data point and then I place my hyperplane in the middle, so it is equidistant from the other two hyperplanes. Then my hyperplane is the one with the maximum margin.

But I still don't understand why it is okay to just say that d = 1.

$\endgroup$

1 Answer 1

0
$\begingroup$

By definition $y$ is labeled as $\{-1, 1\}$. A separating hyperplane is one such that

$\vec{w}^T\vec{x} + b - y > 0$ when $y = 1$

and

$\vec{w}^T\vec{x} + b - y < 0$ when $y = -1$

The above inequalities become equalities at the margins of a maximal margin classifier "band", which is what you have shown above. Which class is labeled as $-1$ and which as $1$ is chosen arbitrary here.

$\endgroup$
1
  • $\begingroup$ hang on a second... why -y? It though it was just wx+b > 0?! $\endgroup$ Oct 21, 2018 at 14:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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