A PROBLEM IN STATISTICAL CLASSIFICATION THEORY

Van Cutsem’s original question was motivated by the following problem: Classification programmes in statistics build classification trees, usually proceeding by successive aggregations of closest neighbours amongst existing classes. How can we measure the way a classification carries useful information and not just "random noise"? Certainly, "good" classification trees should exhibit characteristics that depart significantly from random ones. Hence the need to simulate and analyse parameters of random classification trees.