A multi-objective genetic programming approach to developing Pareto optimal decision trees

Classification is a frequently encountered data mining problem. Decision tree techniques have been widely used to build classification models as such models closely resemble human reasoning and are easy to understand. Many real-world classification problems are cost-sensitive, meaning that different types of misclassification errors are not equally costly. Since different decision trees may excel under different cost settings, a set of non-dominated decision trees should be developed and presented to the decision maker for consideration, if the costs of different types of misclassification errors are not precisely determined. This paper proposes a multi-objective genetic programming approach to developing such alternative Pareto optimal decision trees. It also allows the decision maker to specify partial preferences on the conflicting objectives, such as false negative vs. false positive, sensitivity vs. specificity, and recall vs. precision, to further reduce the number of alternative solutions. A diabetes prediction problem and a credit card application approval problem are used to illustrate the application of the proposed approach.