A feature map sends each data point from its original space into another, usually higher-dimensional, space. Adding a carefully chosen column can turn a curved separation problem into a linear one: (x1, x2) becomes (x1, x2, x1x2), where a plane can separate the points. section 11.3
The transformation supports nonlinear classification by:
- Adding polynomial features such as x1², x1x2, and x2². section 11.3
- Adding similarity columns based on distances to reference points, as with an RBF feature map. section 11.3
- Training a linear classifier in the expanded space, then viewing its projected boundary in the original space. section 11.3
From a curved boundary to a plane
For a circular dataset, add x1² + x2² as a third coordinate. The raised points can be separated by a plane; projecting that plane back produces a circular classifier boundary.







