Computational Complexity Reduction and Interpretability Improvement of Distance-Based Decision Trees

摘要

Classical decision trees proved to be very good induction systems providing accurate prediction and rule based representation. However, in some areas the application of the classical decision trees is limited and more advanced and more complex trees have to be used. One of the examples of such trees are distance based trees, where a node function (test) is defined by a prototype, distance measure and threshold. Such trees can be easily obtained from classical decision trees by initial data preprocessing. However, this solution dramatically increases computational complexity of the tree. This paper presents a clustering based approach to computational complexity reduction. It also discusses aspects of interpretation of the obtained prototype-threshold rules.