Author(s)

Joachim P.P. Costa

Abstract

In this work, we consider the case where we have categorical attributes with a huge number of values in a prediction context. In particular, the method- ology introduced here concerns the use of these attributes in binary decision trees; nevertheless, it is applicable to other prediction methods. The main idea consists in extracting ("mining") the most predictive binary attributes, from the set of initial attributes ("mine"). In order to do this, we consider three different operations: hierarchical clustering, multiplication, and factorization. The first operation, hierarchical clustering, serves for reducing the number of values of a categorical attribute. In fact, by ap- plying the hierarchical clustering method AVL [6, 7] to the set of values of a categorical attribute, we can group these values into clusters. Then, we can

Keywords