Improvements in principal feature classification

We describe algorithms for rapid training and accurate generalization of principal feature classification (PFC) [Q. Li and D.W. Tufts, 1997], an elegant decision tree method with performance comparable to neural networks and popular decision tree methods like CART and C4.5. The improved training method is guaranteed to converge and generates a fully grown decision tree. It accomplishes this with a decrease in complexity and is described in detail to facilitate easy implementation. A generalization methodology based on nonparametric tolerance regions is also developed. The method provides accurate generalization of the fully grown tree and is verified using simulated data