Comparative Analysis of Machine Learning Methods Application for Financial Fraud Detection

This paper addresses the fraud detection problem in the context of Big Data used in remote banking systems. The paper aims to propose a new algorithm for automatic detection of fraudulent transactions using machine learning with a performance that allows to apply it in big data systems. The article identifies promising directions for optimizing the operation of methods for fraudulent transactions detection in anti-fraud systems. Architectural approaches to the operation of anti-fraud systems have been studied. Based on this, an architecture for illegal actions prediction in a near real-time mode was proposed. The research task of the article is to find the most suitable machine learning algorithm, with the least training and prediction time, demonstrating high classification performance. To achieve this goal, an analysis of the supervised and ensemble machine learning algorithms was made. The dataset was preprocessed for the experiment with SMOTE resampling and robust scaling techniques. The chosen methods were compared using different metrics: $F$1 score, AUC and time consumption for training and classification. As a result of a metrics comparison, it was found that multilayer perceptron (MLP) and boosting methods (Adaptive, Gradient, XGBoost) has the highest classification, but MLP outperforms boosting methods in terms of time consumption for classification. Thus, MLP was selected as the most appropriate algorithm for further integration to proposed Big Data architecture. Based on the data obtained during the experiments, the degree of their implementation in fraud detection systems was assessed and architecture for the anti-fraud detection system for big data was proposed.