Analisis perbandingan algoritma Naïve Bayes, k-Nearest Neighbor dan Neural Network untuk permasalahan class-imbalanced data pada kasus credit card fraud dataset

Abstract

The high public interest in transactions using credit cards in the banking sector has the potential for higher credit card fraud. This study uses a credit card fraud dataset that consisting of 284,807 data obtained from Kaggle. The dataset in this study is class-imbalanced data with a comparison between the major class of 99.8% and the minor class of 0.2%. This class-imbalanced data problem will be solved by applying undersampling. In order to determine the performance of the classification algorithm that is most suitable for solving class-imbalanced data problems, a comparison of the Naïve Bayes, k-Nearest Neighbor (kNN) and Neural Network algorithms will be carried out. The t-test in this study was conducted to determine the significance of differences between algorithms. Algorithm performance evaluation uses accuracy and AUC (area under the curve) values. The test results in this study is Neural Network has better performance than other algorithms because it has the highest accuracy value of 93.59% and AUC value of 0.977. Based on the t-test results, the Neural Network with k-NN has a significant difference, in contrast to the Neural Network with Naïve Bayes there is no significant difference.

Fraud Detection Using Random Forest Classifier, Logistic Regression, and Gradient Boosting Classifier Algorithms on Credit Cards

“…This study confirms from previous research which explains that in solving credit card fraud classification problems the algorithm that has the best performance is the Neural Network algorithm with an accuracy value of 93.59% and an AUC score of 0.977. [15] Thus, in this study, it is explained that the results of research using the Random Forest Classifier (RFC) algorithm provide higher accuracy results than previous studies using the Neural Network algorithm with an accuracy value on the data-train percentage of 100% and data-test of 99.99% and the evaluation of the AUC score as the result of testing the algorithm is 0.999 so that it has a better performance effectiveness on the credit card fraud dataset.…”