Model performance comparison and evaluation of bank fraud prevention in simulated datasets

Abstract

The significance of robust fraud detection systems in the banking sector has grown imperative due to the growing prevalence of online transactions. However, the datasets in these particular areas exhibit a greater abundance and diversity. The large amount and variety of data in these domains necessitate the utilization of synthetic sales data, which is derived from real data, as an innovative approach for studying fraud prevention. This study initially derives importance scores for various features through the utilization of random forests. Subsequently, four features that exhibit the highest correlation with fraudulent transactions are selected for further investigation. The training and prediction processes for both random forests and decision tree models are then performed. The study compared the performance of random forests and decision tree models in fraud monitoring using four features. The results indicate that random forests outperform decision trees in terms of accuracy, recall, precision, and F1 scores, with improvements of 0.68%, 0.62%, 0.68%, and 0.65% respectively. These findings provide a comprehensive analysis of the performance comparison between random forests and decision tree models in the context of fraud monitoring.