On handling conflicts between rules with numerical features

Abstract

Rule conflicts can arise in machine learning systems that utilise unordered rule sets. A rule conflict is when two or more rules cover the same example but differ in their majority classes. This conflict must be solved before a classification can be made. The standard methods for solving this type of problem are to use naive Bayes to solve the conflict or using the most frequent class (CN2). This paper studies the problem of rule conflicts in the area of numerical features. A novel family of methods, called distance based methods, for solving rule conflicts in continuous domains is presented. An empirical evaluation between a distance based method, CN2 and naive Bayes is made. It is shown that the distance based method significantly outperforms both naive Bayes and CN2.