Support vector machines - separating hyperplane question
-
20-09-2019 - |
Question
From what I've seen, seems like the separation hyperplane must be in the form
x.w + b = 0.
I don't get very well this notation. From what I understand, x.w
is a inner product, so it's result will be a scalar. How can be it that you can represent a hyperplane by a scalar + b? I'm quite confused with this.
Also, even if it was x + b = 0, wouldn't it be of a hyperplane that passes right through the origin? From what I understand a separating hyperplane doesn't always pass through the origin!
Solution
It is the equation of a (hyper)plane using a point and normal vector.
Think of the plane as the set of points P such that the vector passing from P0 to P is perpendicular to the normal
Check out these pages for explanation:
http://mathworld.wolfram.com/Plane.html
http://en.wikipedia.org/wiki/Plane_%28geometry%29#Definition_with_a_point_and_a_normal_vector
OTHER TIPS
Imagine a plane in a 3d coordinate system. To describe it, you need a normal vector N of that plane and the distance D of the plane to the origin. For simplicity, assume the normal vector has unit length. Then the equation for that plane is x.N - D = 0.
Explanation: x.N can be visualized as a projection of x on the normal vector N. The result is the length of vector x parallel to N. If this length equals D, the point x is on the plane.
A definition of dot product (wich is an inner product) is
x . y = |x| * |y| * cos(a)
Where a is the smallest angle between x and y.
It is easy to see that x . y = 0, if a=90 deg (pi rad).
This means that if you have a fixed normal vector w, a hyperplane given by:
x . w = 0
is the set of all points that x can "point at" given that x has to be orthogonal to w.
Now, a hyperplane given by:
x . w + b = 0
is the set of all points that x can "point at" such that x . w is a constant. As x gets longer, |x| increases, the angle, a, has to get closer to 90 deg (pi rad), cos(a) decreases, to produce the same constant result. If you however take x pointing in the exact opposite direction of w, cos(a) = -1 and |x| = b (provided that w is of unit length).
It turns out that the plane given of this set of points is parallell to x . w = 0, and shifted in space the distance -b (in the direction of w) still given that w is of unit length.
This answer is probably not going to help the op, but hopefully someone else will benefit from it.