Detecção de fraudes bancárias utilizando métodos de clustering

Abstract

Financial institutions have extensively used computerized systems for the provision of their services. Therefore, the number of financial solutions through digital means is increasing. The convenience of using these tools brings with it the important issue of data security in these financial transactions. Among these security issues, this paper focuses on the detection of credit card fraud. By using an electronic payment method, such as a credit card, to make purchases or contract services. When the credit card is used, data from that purchase is added to a transaction table, some of which are the purchase moment and their values. These data can be used to construct a typical pattern of behavior. When system security is compromised, such as when the credit card is used by someone other than the user, we have a fraud. The purpose of this project is to detect unusual behaviors, that are, atypical transactions in comparison to others. The data used comes from a real database with history of credit card transactions and anonymous. The database is labeled with 284,807 standard transactions and 492 frauds. In this study, an analysis of the database in question was performed, after this analysis the unnecessary resources were removed. Having the data set ready, a selection was made of the clustering methods that would be applied, and subsequently tested hybrid approaches of these algorithms. As a result, in the best cases, fraud detection rates greater than 88% were obtained, using the clustering methods DBSCAN and K-Means.The DBSCAN achieved better results using Manhattan metric.