Hierarchical Backward Search Method: A New Classification Tree Using Preprocessing by Multilayer Neural Network

Abstract

This paper proposes a novel pattern recognition algorithm. The new algorithm consists of two stages: the pyramid making stage and the classification making stage. The classification tree generated by the algorithm is called “PACT” (Pyramid Architecture Classification Tree), which is just a kind of classification tree but has a great ability to deal with huge dimensional real scale problem.

PACT does not only shows good performance but also induces a new idea: a fractal characteristic which a category set shows in huge dimensional feature space. The new idea accounts for how awkward conventional algorithms, including neural networks and classification trees, are when applied to huge dimensional pattern classification problems. The new idea gives a new positive reason for hierarchical pattern processing.