Support vector machines

1 min read

Built upon Simple Intuitive maximal margin classifier

MMC requires classes to be separable by a linear boundary

SVC solves a broader case

SVC uses a soft margin to solve inseparable cases

a few samples is on the incorrect side, but the general public is better

SVM is a further extension and can solve nonlinear boundaries

SVM is able to create nonlinear boundaries

One way is enlarging the feature space

A better way is using a kernel

A kernel is simply a function

Instead of the inner product of features, we apply a function to it, and call this function a kernel

This essentially means fitting an SVC in a higher dimensional space

Using a kernel is computationally efficient

SVM is connected to logistic regression, their loss functions are very similar

Also kernels can be used with other methods, too. Although they are used with SVMs mostly

You can use SVMs for more than 2 classes.

  • one vs one Create SVMs for all 2-pairs of classes. In the end a sample belongs to the class to which it is most frequently assigned

  • one vs all

January 16, 2020