An AutoEncoder enhanced light gradient boosting machine method for credit card fraud detection.

Area of Science:

• Machine Learning
• Data Science
• Financial Technology

Background:

• Online financial transactions are increasingly vulnerable to fraud.
• Existing machine learning methods struggle with large, imbalanced financial datasets.
• Traditional evaluation metrics are insufficient for imbalanced anomaly detection.

Purpose of the Study:

• To propose a robust AutoEncoder enhanced LightGBM method for credit card fraud detection.
• To address the challenges of high dimensionality and class imbalance in financial data.
• To improve the accuracy and efficiency of anomaly detection in financial transactions.

Main Methods:

• Utilized an AutoEncoder for feature reconstruction to handle high dimensionality.
• Integrated the LightGBM algorithm, an enhancement of Gradient Boosting Decision Tree (GBDT), for efficient anomaly detection.
• Employed advanced evaluation metrics like AUC, MCC, and BCR alongside traditional ones.
• Investigated the combination of the proposed model with sampling techniques, notably SMOTE.

Main Results:

• The AutoEncoder-LightGBM model significantly outperformed existing methods on two financial datasets.
• Achieved a recall of 94.85% and BCR of 97% on a credit card fraud dataset.
• Demonstrated superior performance on a high-dimensional Santander bank transaction dataset, improving recall and F-measure.
• The combination with SMOTE yielded the best results, with AUC 96.83% and F-measure 80.27% on the fraud dataset.

Conclusions:

• The proposed AutoEncoder enhanced LightGBM model offers a robust and superior solution for credit card fraud detection.
• The method effectively handles large, high-dimensional, and imbalanced financial datasets.
• Combining the model with the SMOTE algorithm further enhances its performance, proving effective for anomaly detection.