A Technique For Analyzing Banking Transactions To Identify Fraudulent Activities In E-commerce

Abstract

The article presents a technique for analyzing banking transactions using machine learning methods and graph algorithms to identify fraudulent activities in e-commerce. The first stage of the proposed technique involves training an ensemble of classifiers on a labeled transaction data set. The second stage of the methodology involves launching the Personalized Page Rank (PPR) algorithm on the constructed directed transaction graph to search for counterparties who interact with fraudulent or affected by fraudulent counterparties more than others. Interpreting the PPR results allows one to identify potential victims of fraudsters. An approach to identify fraudulent schemes for cashing out of funds through individual entrepreneurs is also proposed, based on the search in large graphs for sub graphs satisfying a given property. The application of the technique is illustrated on a publicly available dataset of transactions of European bank card holders. The estimates of the work of the classifiers (precision, completeness, F -measure) are given. To build a transaction graph, the dataset was artificially extended with the payer and recipient fields. The results of the search for potential victims of fraud using the PPR algorithm are discussed, as well as the approach for construction of queries in Apache Spark to search for cash-out chains of a given length.