Unsupervised splitting rules for neural tree classifiers
Michael Perrone, Nathan Intrator · 2003
The authors present two unsupervised neural network splitting rules for use with CART-like neural tree algorithms in high-dimensional data space. These splitting rules use an adaptive variance estimate to avoid some possible local minima which arise in unsupervised methods. They explain when the unsupervised splitting rules outperform supervised neural network splitting rules and when the unsupervised splitting rules outperform the standard node impurity splitting rules of CART. Using these unsupervised splitting rules leads to a nonparametric classifier for high-dimensional space that extracts local features in an optimized way.>