Financial Fraud Detection Using Isolation Forest And DBSCAN Clustering Techniques

Abstract

Financial fraud is an urgent problem in modern banks and fintech companies that leads not only to significant financial losses but also undermines customer trust in these companies. Rule-based methods and supervised learning (SL) models have some limitations in recognizing new cases and infrequent incidents. Hence, the research aimed to develop a hybrid architecture based on Isolation Forest (IF) and DBSCAN algorithms to improve the efficiency of fraud detection while providing the interpretability of results. The model was tested using the Kaggle Credit Card Fraud dataset that included 284,807 entries with highly unbalanced classes (fraudulent cases = 492). The preprocessing step included normalization of the Time and Amount variables. Initially, IF identified outliers. The next step was to apply DBSCAN clustering using those anomalies. Different measures were used for performance analysis of the proposed approach. The values of 99.88% for accuracy, precision 0.93, recall 0.87, F1 score 0.90, and ROC-AUC score 0.96 were achieved using the hybrid model. From the cluster evaluation results, the former clusters are clearly separated with Silhouette score being 0.61 and the DBI value of 0.45. It is evident that the hybrid approach has produced understandable results which can be used for fraud detection and grouping. Through comparative study, it is clear that the hybrid approach outperforms the IF. It shows that combination of IF and DBSCAN will be effective for fraud detection due to interpretability.