Should decision trees be learned from examples or from decision rules?

Abstract

A standard method for determining decision trees is to learn them from examples. A disadvantage of this approach is that once a decision tree is learned, it is difficult to modify it to suit different decision making situations. An attractive approach that avoids this problem is to learn and store knowledge in a declarative form, e.g., as decision rules, and then, whenever needed, generate from it a decision free that is most suitable in any given situation. This paper describes an efficient method for this purpose, called AQDT-1, which takes decision rules generated by the learning system AQ15 and builds from them a decision tree optimized according to a given quality criterion. The method is able to build conventional decision trees, as well as the so-called “skip noder” trees, in which measuring attributes assigned to some nodes may be avoided. It is shown that “skip-node” trees can be significantly simpler than conventional ones. In the experiments comparing AQDT-1 with C4.5, the former outperformed the latter both in terms of the predictive accuracy as well as the simplicity of the generated decision trees.