Volume 10, Issue 1

Articles
Trott's Corner
New Products
New Publications
Calendar
News Bulletins
New Resources
Classifieds

Editorial Policy
Staff and Contributors
Submissions
Subscriptions
Back Issues
Contact Information

A Flexible Implementation for Support Vector Machines

Regression Analysis with SVMs

So far we have considered SVMs as a tool for pattern recognition only. It is also possible to use the SVM framework for regression problems. Consider a function to be approximated; for example, a quadratic.

We can adapt the SVM method to the regression setting by using a -insensitive loss function

where is the SVM approximation to the regression function . This loss function determines how much a deviation from the true is penalized; for deviations less than , no penalty is incurred. Here is what the loss function looks like.

Using this idea, the regression problem is transformed to a classification problem: any such that may be considered "correctly classified." MathSVM solves such problems using the RegressionSVM function, parameterized by and a penalty constant . Here we again try a polynomial kernel.

The function RegressionSVMPlot provides convenient plotting of the resulting regression function. As with SVMPlot, the kernel type used is supplied as a parameter. Note how support vectors in this case are chosen as the data points that are furthest away from the regression line.

We can, of course, also obtain the analytical expression of the estimated regression function.

Two-Dimensional Example

We can use SVM regression with domains of any dimension (that is the main advantage). Here is a simple two-dimensional example.

Here is the regression function.

There are no specialized 3D plots for regression in the MathSVM package. Here is the usual Plot3D visualization.