Fraudulent Transaction Detection in FinTech using Machine Learning Algorithms

Abstract

With the advancement of e-commerce, online transactions purchases using credit and debit cards have drastically increased. This has caused a burst in credit and debit card fraud and has become a profoundly significant global issue. Fraud touches every area of our lives and is a growing concern that effects both businesses and customers. As machine learning techniques provide unique and efficient solutions, they are applicable in various types of problems. Recently, machine learning algorithms have been widely applied as a data mining technique for classification problems. In this paper, a binary classification problem is considered where a transaction can be classified as either fraudulent or legitimate transaction. The goal is to classify the transactions using five different machine learning algorithms. The transaction dataset (Task1 and Task2) is preprocessed, and then SGD, DT, RF, J48 and IBk machine learning classifiers are applied. After applying the classifiers, the results are compared to analyze which classifier performs the best. Based on the experimental results, it is found that the accuracy percentage of all the five classifiers for Task1 and Task2 datasets is ranging between 97.78% to 98.1%, with no major difference. As the dataset is highly imbalanced, the kappa statistic value is also considered. For both datasets, the RF classifier had the greatest value of kappa statistics, whereas SGD and J48 had the lowest value for Task1 and Task2, respectively. Other evaluation metrics were also considered for evaluating the performance of the applied classifiers. Overall, these classifiers achieved similar results for Task2 dataset. As negative Kappa statics and MCC values were obtained, the SGD classifier for Task2 dataset had the worst results in comparison. Based on evaluation criteria such as Kappa statistic and MCC values RF outperformed the others for both the datasets.