Machine Learning Based Clustering Algorithms and Graph Analysis as Approach for Improving Anti-fraud Model in Credit Institutions

In recent years, the development and implementation of new technologies in banking services has been proceeding at extremely fast pace, with both positive (for example, faster and easy accessible online services) and negative effects (the scale of fraud has increased significantly). Therefore, forward-thinking organizations keep exploring ways to increase the efficiency of anti-fraud model they use. Among the most effective tools stands out machine learning methods. This paper explores the possibility of improving the existing anti-fraud model used in real banking practice. The aim of this work is to improve the quality of the above mentioned model to more accurately identify fraud. To achieve the purpose of the work, for a start, loan applications were clustered on a machine learning based methods. The following clustering methods were used: k-means, k-prototypes and DBSCAN. A comparison between these methods was made with respect to the corresponding metrics: Elbow, Silhouettes and IV (Information Value). Then it was investigated how the resulting clusters affect existing anti-fraud model score. Then, the analysis of the graphs based on data obtained from applications was carried out. The following metrics, calculated on the basis of social graphs, were used: centrality metrics (degree centrality, betweenness centrality, closeness centrality), PageRank, cluster number and size, shortest paths to the fraud essence. Both of the above mentioned approaches for improving the model were compared to each other. The work resulted in some improvement of the original anti-fraud model. Keywords: fraud, anti-fraud model, loan applications, machine learning, clustering, k-means, k-prototypes, DBSCAN, graph, centrality metrics, PageRank, shortest paths to the fraud essence.