The soft-margin support vector machine (SVM) is a classic method for supervised learning.
It looks to find a large-margin classifier: one for which its decision boundary is far from the examplars. This is unlike the linear perceptron, which will not in general guarantee large margins. The SVM model minimizes the following loss:
\[L(w) = \lambda ||w||^2 + \sum_{(x, y)} L_H(w, x, y)\]Here, $H_L$ denotes the hinge loss:
\[L_H(w, x, y) = \left \{ \begin{array}{rl} 1 - y \langle w, x \rangle &, \textrm{if} \ y \langle w, x \rangle \le 1 \\ 0&, \textrm{otherwise} \end{array} \right .\]The hinge loss is particularly useful for classification because of two reasons. First, it is convex, which means that there exist algorithms that minimize $L(w)$ efficiently. Second, in the region where the misclassification loss (of a linear classifier) returns zero, the hinge loss has compact support, specifically in a way such that for points that are sufficiently far from the decision boundary, the hinge loss is zero.
Crucially, that second reason implies that the position of correctly classified points that are sufficiently away from the decision boundary does not matter for the classifier. One way to see this intuitively is that if you are given a classifier that attains the minimum, together with one point for which this classifier gives a hinge loss of zero, then if you wiggle this point in any direction, the loss will still be zero, and that means that the classifier will still be optimal, even with this wiggled point. (Of course, if you wiggle the point so much that it crosses into the region of the hinge loss where the value is non-zero, then you’ll potentially change the classifier.)
As a result, after the training procedure finishes, we can identify the subset of input points which influence the decision: these are the support vectors (they “support” the decision). Support vectors are particularly important in the “kernelized” formulation of the SVM. A general kernelized linear classifier needs to potentially access all training points to make a test-time prediction. The SVM, in contrast, needs only to store the support vectors, which might be a small fraction of all the input points.
You can see the support vectors in the example below, where we train a support vector machine with 2D data (using a quadratic polynomial kernel). As $\lambda$ gets smaller, the penalty for making hinge loss errors gets comparatively larger, so the margin of the classifier itself gets smaller. As a result, the number of support vectors gets smaller. Note how the support vectors are all points that are either misclassified, or are points inside the $[-1, 1]$ range of the classification values: “within the margin”.