Improving Imbalanced Data Classification Using Deep Learning

Abstract

Classifying imbalanced data is a difficult task in many machine learning applications, especially in the context of fraud detection. This paper evaluated the performance of traditional models (e.g., Random Forests, XGBoost, and CatBoost) against the performance of deep learning models. While the traditional models were able to obtain high accuracy, they struggled to identify the rare classes (i.e., fraudulent transactions) when the F1 scores did not get above 0.33. In turn, a deep learning model was proposed that applied ideas such as class weights, decision thresholds, and F1-maximizing training objectives and was designed to employ voting of multiple submodels. The results demonstrated that the proposed model (Ensemble Neural Network) was able to achieve an F1 score of 0.5997 and an AUC-PR score of 0.6205 which outperformed the traditional methods previously used in the study. This design was used to achieve a better balance between identifying the rare classes and overall model performance.