Improving Money Laundering Detection Using Optimized Support Vector Machine

Abstract

Identification of financial transactions as suspicious or fraudulent transactions that are indicated as money laundering is mostly done manually so that it is not optimal. Data mining techniques can be a solution to overcome the limitations of the manual method. The main challenge in applying data mining techniques for financial fraud detection is an imbalanced dataset, where the proportion of fraud class is much smaller than non-fraud. This causes the model to produce unbalanced precision and recall, resulting in a low f1score. It means that the model can predict one class well, but not with another class. In this paper, the approach to fraud detection in financial transactions is carried out with classifier optimization based on Support Vector Machine (SVM). Optimization is performed by tuning the kernels and hyperparameters combined with the Random Under Sampling (RUS) technique. Specifically, RUS is used to handle imbalanced datasets and cut model training time. With this combination technique, the classifier can detect fraud more effectively with an increase in precision of 40.82% and f1-score of 22.79% compared to the previous study. A combination technique can be an approach to cover weaknesses left behind by a single method.