Ensemble learning based predictive modelling on a highly imbalanced multiclass data

Abstract

Class imbalance in the real-world datasets is a big challenge and the domains such as fraud detection, calamity occurrences, bankruptcy prediction etc. are prone to class imbalance due to the nature of occurrences of the events. In this paper, the detailed research using six ensemble machine learning techniques is applied to the undersampled, oversampled and the original dataset and the results are compared. The results of the research study indicates that amongst the applied six ensemble learners, the best learner is Random Forest algorithm (with entropy gain) implemented using ten-fold cross validation on the SMOTE oversampled dataset. 0.95 AUC and 0.8689 accuracy i.e. an increase of 4% in accuracy and substantial increase in other performance indicators is observed as compared to the remaining five ensemble classifiers.