Genetic Algorithm-Based Oversampling Technique to Learn from Imbalanced Data

Abstract

Availability of data from many different applications such as surveillance systems, security appliances, finances has been continuously expanding. Many machine learning (ML) and data mining models have shown promising power in learning from the available data. However, the problem of learning an ML classifier from imbalanced data is still a challenging problem. This problem is often regarded as the imbalanced learning problem. In this problem, there is more amount of information known from the majority classes than the minority classes. In such a learning environment, the classifier during training over-fits to the former classes and under-fits to the minority classes. Distance-based strategy, for example, SMOTE, has been quite useful to oversample the minority classes that essentially uses nearest neighbor samples from the available samples. In this paper, we propose a notion of employing genetic algorithm (GA) that would essentially learn the probability distribution from the available data to generate the minority class samples for binary classification problems. We validate and test our proposed oversampling strategy by training three different kinds of classifiers. The comparative analysis with SMOTE-based oversampling and the proposed GA-based oversampling shows promising results for a selected ten very popular imbalanced datasets.