A distribution-preserving method for resampling combined with LightGBM-LSTM for sequence-wise fraud detection in credit card transactions

Abstract

Fraud detection is a challenging task that can be difficult to carry out. To address these challenges, a comprehensive framework has been developed which includes a new resampling method combined with a data-dependent classifier that can detect fraud effectively. The proposed framework uses two hybrid approaches that leverage the strengths of a One-Class Support Vector Machine (OCSVM) with the Synthetic Minority Oversampling Technique (SMOTE) and random undersampling. The distribution of fraud instances is effectively preserved by this innovative framework. The comparison of the probability functions of fraud data before and after resampling is demonstrated, indeed. Afterward, The outputs of our hybrid approaches are analyzed using two distinct models, the Light Gradient-Boosting Machine (LightGBM) and the Long Short-Term Memory (LSTM) model. Our case study on European credit cards has consistently demonstrated the effectiveness of our techniques over existing methods, achieving a high F1 score of 87% with a corresponding AUC score of 96% in non-sequential fraud detection and The F1 score of 85% with an AUC score of 87% in sequential fraud detection. Additionally, we have developed an innovative algorithm for determining optimal window sizes for sequence-wise fraud analysis, which recommends window sizes of 3 for the European dataset, highlighting the efficacy of sequence-wise analysis. Overall, the proposed framework, not only offers a promising solution to enhance fraud detection accuracy, but it also reduces false positives.