Credit Card Fraud Detection with Data Sampling
T. John Berkmans, S. Karthick · 2022
Credit card fraud, although not directly impacting banks, does have repercussions for the financial sector as a whole. Criminals pose a significant threat to safety since they are continually thinking up new methods to commit these types of fraud. Early detection of fraudulent behavior is therefore vital to retain customer trust and defend the firm. Since lawful transactions far outnumber fraudulent transactions, which typically make up less than 1% of all transactions, addressing the class imbalance issue in the data presents a significant challenge for developing fraud detection algorithms. It is challenging to identify a positive example (fraudulent case), and this challenge increases as more data is gathered, leading to a lower proportion of positive examples. Because of this, research is very important. Models for making predictions were trained in this study utilizing a variety of sampling strategies, including the ANN, GBM, and the RF. Models used SMOTE, RUS, DBSMOTE, and SMOTE plus ENS were all examples of Synthetic Minority Over-Sampling Technique (SMOTEENN). This research suggests that SMOTE-based sampling techniques will provide positive results. The highest recall (0.81) was achieved by the SMOTE sampling method when a DRF classifier was used. It was determined that this classifier has an accuracy score of 0.87. The Stacked Ensembling algorithm was accomplished using altogether of collected data, and its average performance was 0.78, making it the clear winner. As a fraud detection model, the Stacked Ensemble has performed well in most sampling operations.