Piecewise Linear Classification with Hyperplanes

Figure1 - Piecewise linear classification of points with hyperplanes
Any approach to point classification is a compromise between minimizing the error on test data sets and maximizing the probability that the approach will perform well on new data. Reconciling these two conflicting goals is the motivation behind the algorithm presented in the article "Piecewise Linear Classifiers with an Appropriate Number of Hyperplanes" by H. Tenmoto, M. Kudo and M. Shimbo [1]. They present a method for constructing a piecewise linear classifier using a minimal number of hyperplanes, based on a maximum classification error tolerance.

This report consists of a general overview of classification, a description of the classification algorithm by H. Tenmoto et al., followed by a discussion of some of the merits and drawbacks of the algorithm. An interactive Java applet is also provided to demonstrate the algorithm.

This web page prepared by
Matt Toews (mtoews@cim.mcgill.ca)
as a term project for the course
Computer Science 644 - Pattern Recognition
at the
McGill Center for Intelligent Machines