Estimation of optimal threshold shifting to handle class imbalance in credit card fraud detection using machine learning techniques.

Class imbalance is one of the major challenges faced by the machine learning research community. The imbalance in the data degrades the performance of the model. The widely used resampling methods can miss important instances during undersampling. On the other hand, oversampling can cause bias in the data that lead to overfitting of the model. To avoid this problem an alternate solution of estimating the threshold for an imbalanced dataset was adapted. In general, the default threshold is fixed to 0.5, which is not suitable for imbalanced datasets. The optimal threshold is estimated by maximizing the f1-score and Geometric mean (G-mean), the two widely used metrics for imbalanced learning. The proposed approach preserves the original data without any alteration. To compare the performance six ensemble classifiers such as Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LGBM), eXtreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Random Forest (RF) and Classic Gradient Boosting (CatBoost) are selected. The IEEE-CIS fraud detection dataset from Kaggle having an imbalance ratio of 27.58 was used for this study. Experimental results show that the precision, recall and f1-score are improved by shifting the threshold values for binary classification. We observed that the performance of XGBoost (0.7325) and LGBM (0.7123) was comparably high in terms of f1-score. Hence, the selection of optimal threshold helps to provide betterfindings. [ABSTRACT FROM AUTHOR]