Dealing with Class Imbalances for Detection of Fraudulent Credit Card Transactions

Abstract

Data Science is a multidisciplinary field that uses algorithms to extract knowledge from unstructured and structured data. The major problem that was faced when dealing with this dataset is class imbalances. The dataset consists of two classes, fraud, non-fraud transactions. The ratio of fraud to non-fraud transactions is 1:500 in the dataset, meaning that there could be potential overfitting towards the non-fraudulent transactions resulting in faulty predictions. The proposed model for overcoming this challenge is by applying SMOTE (Synthetic Minority Oversampling Technique), which helps in overcoming the problem of class imbalances. As a result of all the processing, it has been found that LightGBM after the application of SMOTE produces an accuracy of 99.2% offering the best result amongst the other models used for surveying purposes.