Comparison of Efficiency for Imbalanced Data Classification via Simulation

The aims of this research are to compare the efficiency of imbalance techniques between over sampling and hybrid methods and to compare performance of classification techniques: random forest, logistic regression, and support vector machine, via simulation. The study is given by high imbalanced data and predicted variables which are mostly categorical data. The criteria of the simulation are sample sizes, ratio of the number of predicted variables between categorical variables and continuous variables, and odds ratio. The results shown that balancing data with over sampling method before classify had higher accuracy, sensitivity, and specificity than hybrid method in each sample sizes. In addition, the balanced data classified with random forest had the highest accuracy, sensitivity, and specificity, the average were 0.996, 0.999 and 0.998 respectively. Moreover, logistic regression technique yields less accurate classification when the number of categorical variables is higher. The result of research can be used as a guideline for choosing a data balancing method which appropriate to data conditions in real situations.