Alvin Quashie

Machine Learning (ML) is a branch of Artificial Intelligence (AI) that allows computers to learn from  previous knowledge (data) and improve their predicting abilities without being specifically programmed to do so.

This research focuses on the application of the following supervised ML algorithms for credit card fraud detection: Decision Tree, Random Forest, Artifcial Neural Network, Naive Bayes  and Logistic Regression. In this work, a credit card fraud dataset generated from European credit cardholders is utilized

The experiments were carried out in two folds. In the first step, a classifcation process was conducted using F = {v1, v2, v3, v4, v5}. For each feature vector in F, the following methods were trained and tested: RF, DT, ANN, NB and LR. The results are depicted in Tables 2, 3, 4, 5, 6. As shown in Table 2, both the ANN and the RF algorithms obtained the highest test accuracy (TAC) of 99.94% using v1. However, the RF method obtained the best results in terms of precision. In Table 3, the results that were obtained using v2 demonstrate that the best model is the RF approach with an accuracy of 99.93%. In Table 4, the RF method also obtained the best fraud detection accuracy of 99.94% using v3. Table 5 presents the results that were achieved by v4 whereby the DT obtained an accuracy of 99.1% and a precision of 81.17%. Table  6 depicts the outcomes that were obtained when using v5. In this case, the RF attained a fraud detection accuracy of 99.98% and precision of 95.34%. In comparison to the results obtained by v1, v2, v3 and v4; v5 obtained the best results. Moreover, looking at the outcomes presented in Tables 2, 3, 4, 5, 6, the NB method under performed in terms of Recall, Precision and F1-Score.

In this research, a GA based feature selection method in conjunction with the RF, DT, ANN, NB, and LR was proposed. Te GA was implemented with the RF in its ftness function. Te GA was further applied to the European cardholders credit card transactions dataset and 5 optimal feature vectors were generated. Te experimental results that were achieved using the GA selected attributes demonstrated that the GA-RF (using v5) achieved an overall optimal accuracy of 99.98%. Furthermore, other classifers such as the GA-DT achieved a remarkable accuracy of 99.92% using v1. Te results obtained in this research were superior to those achieved by existing methods. Moreover, we implemented our proposed framework on a synthetic credit card fraud dataset to validate the results that were obtained on the European credit card fraud dataset. Te experimental outcomes showed that the GA-DT obtained an AUC of 1 and an accuracy of 100%. Seconded by the GA-ANN with an AUC of 0.94 and an accuracy of 100%. In future works, we intend to use more datasets to validate our framework.