Tree classifier design with a permutation statistic

摘要

This paper proposes a permutation statistic as a basis for designing binary tree classifiers of d-dimensional patterns. The statistic permits an objective and easily interpreted evaluation of a (feature, threshold) pair that is the core of the design paradigm. It has a known distribution that is independent of the number of training patterns. After defining the permutation statistic and explaining its properties, the paper proposes the design procedure whereby each non-terminal node is assigned a feature and a threshold on that feature. The design procedure is demonstrated on six test sets and compared to designs based on mutual information.