Evaluation of the Factors Affecting Classification Performance in Class Imbalance Problem

In binary classification, when the distribution of numbers in the class is imbalanced, we are aimed to increase the accuracy of classification in classification methods. In our study, simulated data sets and actual data sets are used. In the simulation, the "BinNor" package in the R project, which produces both numerical and categorical data, was utilized. When simulation work is planned, three different effects are considered which may affect the classification performance. These are: sample size, correlation structure and class imbalance rates. Scenarios were created by considering these effects. Each scenario was repeated 1000 times and 10-fold cross-validation was applied. CART, SVM and RF methods have been used in the classification of data sets obtained from both simulation and actual data sets. SMOTE, SMOTEBoost and RUSBoost were used to decrease or completely remove the imbalance of the data before the classification methods were applied. Specificity, sensitivity, balanced accuracy and F-measure were used as performance measures. The simulation results: the imbalance rate increases from 10 to 30, the effect of the 3 algorithms on the classification methods is similar accuracy. Because the class imbalance has become balanced