Date of Award

1-2023

Document Type

Thesis

Degree Name

Master of Science (MS)

College/School

College of Science and Mathematics

Department/Program

Computer Science

Thesis Sponsor/Dissertation Chair/Project Chair

Michelle Zhu

Committee Member

Jiayin Wang

Committee Member

Rui Li

Abstract

Credit card fraudulent transactions are becoming an ever-growing problem in the financial market. There has been a rapid increase in the rate of fraudulent transactional activities in recent years producing considerable financial loss to many companies, organizations, and government agencies. These numbers are anticipated to increase in the near future and many scholars in this field are focused on detecting fraudulent transactions early on, using advanced Machine Learning techniques. However, credit card fraud transaction detection is not easy for two reasons: (I) fraudulent methods usually vary for each attempt, and (II) the dataset is extremely imbalanced with many more normal transactions than fraudulent cases. A random oversampling approach is used to solve this issue by utilizing precision, recall, F-beta scores, and thresholds. The original dataset contains 284,807 transactions with only 492 transactions labeled as fraudulent. Predictive models, such as logistic regression, random forest, decision tree, and Recurrent ANN with dense layers and sequential model, Adam optimizer, and cross-entropy are used, in combination with different resampling methods. The model performance is evaluated using metrics, including recall, precision, F1-score, ROC AUC score, and accuracy for fraud cases. Our experimental results indicate that ANN outperformed other models. Our ANN utilizes dense layers and some sequential models with optimizer and loss type. The ANN approach, combined with different resampling methods, has been implemented to predict the nature of a certain credit card transaction. To further enhance the performance, the Long Short Term Memory model (LSTM) with multilayer perception is introduced. Multilayer Recurrent Neural Network based LSTM performs better than artificial neural network-based models and improves overall model accuracy. It also improves the accuracy of fraudulent case prediction. We further explore RNN-LSTM and conduct experiments by introducing bi-directional LSTM. This Bi-directional LSTM improves the accuracy of the model to 99.97%. We also find out that some data fields of the credit card transaction have a higher contribution to the detection.

File Format

PDF

Share

COinS