Kybernetika 34 no. 4, 479-484, 1998

Piecewise linear classifiers preserving high local recognition rates

Hiroshi Tenmoto, Mineichi Kudo and Masaru Shimbo

Abstract:

We propose a new method to construct piecewise linear classifiers. This method constructs hyperplanes of a piecewise linear classifier so as to keep the correct recognition rate over a threshold for a training set. The threshold is determined automatically by the MDL (Minimum Description Length) criterion so as to avoid overfitting of the classifier to the training set. The proposed method showed better results in some experiments than a previous method.