Fraud Detection in Financial Transactions Using Machine Learning

Abstract

Financial transaction fraud is an ongoing threat with significant economic loss and on customers' trust. This paper discusses Fraud Detection in detail with machine learning technique on a given data set of a transaction. We investigate the patterns revealed from the users and the transactions in the database when the user is performing fraudulent transactions, and test several classification models that can be used to detect fraud, which includes logistic regression model, random forests, support vector machines, gradient boosting, and neural networks. The performance of the models is investigated in terms of accuracy, precision, recall, F1 score and ROC-AUC metrics. Based on our experiments, the best detection overall performances are obtained for the tree-based ensemble models (Random Forest and XGBoost) with XGBoost getting the most optimum fraud Recall and F1-Score. Through the data analysis results (such as account age, transaction frequency etc.) and the model comparison, we expound an improved method which is based on combining the ensemble of best models with data imbalance countermeasures to increase the recall of fraudulent cases. We also have an end-to-end machine learning pipeline on Python to detect frauds from preprocessing the data, training the models, evaluating them, and deploying for fraud prediction. Also, a literature review of twenty-five recent studies on fraud detection is given, and the algorithms used, datasets and major contributions of these studies were summarized. The textbook ensemble technique, as proposed gives better fraud detection performance as it gains on the order of ~3-5% improvement against the best single model performance on F1-score, with acceptable precision, thereby underscoring the usefulness of hybrid modeling with specialized techniques for this field. The results emphasize that utilizing various models and domain-specific feature engineering can be of great benefit in fraudulent transaction detection, while also neg