Penerapan Modified ADASYN untuk Meningkatkan Akurasi Pendeteksian Pola Fraud pada Transaksi Kartu Kredit

Abstract

One of the most influential factors in accuracy is class imbalance, this is reviewed in a study conducted by Gameng et al. (2019). In the study of Bagga et al. (2020), the Pipelining method combined with ADASYN the accuracy can reach 0.99999. The problem in this study is that accuracy may not necessarily reach 0.99999 if using a dataset outside the dataset they are using and if using a Classificationnn algorithm other than pipelining. In a study conducted by Dornadula & Geetha (2019), the highest accuracy was only 0.9994. In the research conducted by Makki et al. (2019), the Classificationnn model that uses the class balancing method has lower accuracy.

In this thesis, Modified ADASYN is used because in the research of Gameng et al. (2019) its accuracy, precision and f1-score surpassed ADASYN and SMOTE. Pipelining method is used because in the study of Bagga et al. (2020), Pipelining can make Classificationnn accuracy up to 0.99999.

As a result of testing, this thesis concludes that Modified ADASYN has not been able to obtain an accuracy of 0.999999 on two different datasets. In this thesis, Modified ADASYN is able to increase the accuracy of K-NN to 0.9995148 and 0.97617554 using the first and second datasets. Modified ADASYN can outperform SMOTE, ADASYN, One-Class Classificationnn and Cost Sensitive. In this thesis, it is found that the optimal K value in Modified ADASYN can vary depending on many parameters and sample data.