A Logarithmic Distance-Based Multi-Objective Genetic Programming Approach for Classification of Imbalanced Data

Abstract

Standard classification algorithms give biased results when data sets are imbalanced. Genetic Programming, a machine learning algorithm based on the evolution of species in nature, also suffers from the same issue. In this research work, we introduced a logarithmic distance-based multi-objective genetic programming (MOGP) approach for classifying imbalanced data. The proposed approach utilizes the logarithmic value of the distance between predicted and expected values. This logarithmic value for the minority and the majority classes is treated as two separate objectives while learning. In the final generation, the proposed approach generated a Pareto-front of classifiers with a balanced surface representing the majority and the minority class accuracies for binary classification. The primary advantage of the MOGP technique is that it can produce a set of good-performing classifiers in a single experimental execution. Against the MOGP approach, the canonical GP method requires multiple experimental runs and a priori objective-based fitness function. Another benefit of MOGP is that it explicitly includes the learning bias into the algorithms. For evaluation of the proposed approach, we performed extensive experimentation of five imbalanced problems. The proposed approach’s results have proven its superiority over the traditional method, where the minority and majority class accuracies are taken as two separate objectives.