Second Tier for Decision Trees

Abstract

A learner's performance does not rely only on the representation language and on the algorithm inducing a hypothesis in this language. Also the way the induced hypothesis is interpreted for the needs of concept recognition is of interest. A flexible methodology for hypothesis interpretion is offered by the philosophy of a learner's second tier as originally suggested by Michalski (1987). Here, the potential of this general approach is demonstrated in the framework of numeric decision trees. The second tier improves classification performance, increases ability to handle context, and facilitates transfer of a hypothesis between different contexts. 1 Introduction This paper concentrates on concept learning from examples described by vectors of numeric variables and classified as positive and negative instances of the concept. No background knowledge is considered. The learner takes as input a set of pairs [x; c(x)], where x = [x 1 ; x 2 ; : : : ; xn ] is the vector describing the examp...