Self-balancing recursive partitioning algorithm for classification problems

Abstract

Creating an effective classification model has been played an important role in knowledge discovery in a database methodology for the past several years. However, there is a critical issue that significantly affects the classification performance appearing in many real-world situations, which is called a class imbalanced problem. In this dissertation, a classification model built based on the recursive partitioning algorithm is improved under the concept of modified entropy components for handling a classification problem regardless of the class imbalanced situation. Three methodologies are introduced for achieving different purposes. The first methodology is presented to classify a binary-class imbalanced dataset dealing only with numeric attributes, and then it is enhanced to deal with the multi-class case in the second methodology. The third methodology is designed to work with a dataset consisting of both numeric attributes and categorical attributes. The experimental results on both synthetic datasets and real-world datasets from the UCI repository show that these proposed methodologies significantly outperform other existing methods.