Automatic machine learning algorithms for fraud detection in digital payment systems

Oleh Kolodiziev,Aleksey Mints,Olha Lozynska,Inna Pleskun,Pavlo Sidelov

doi:10.15587/1729-4061.2020.212830

Abstract

Data on global financial statistics demonstrate that total losses from fraudulent transactions around the world are constantly growing. The issue of payment fraud will be exacerbated by the digitalization of economic relations, in particular the introduction by banks of the concept of Bank-as-a-Service, which will increase the burden on payment services. The aim of this study is to synthesize effective models for detecting fraud in digital payment systems using automated machine learning and Big Data analysis algorithms. Approaches to expanding the information base to detect fraudulent transactions have been proposed and systematized. The choice of performance metrics for building and comparing models has been substantiated. The use of automatic machine learning algorithms has been proposed to resolve the issue, which makes it possible in a short time to go through a large number of variants of models, their ensembles, and input data sets. As a result, our experiments allowed us to obtain the quality of classification based on the AUC metric at the level of 0.977‒0.982. This exceeds the effectiveness of the classifiers developed by traditional methods, even as the time spent on the synthesis of the models is much less and measured in hours. The models' ensemble has made it possible to detect up to 85.7 % of fraudulent transactions in the sample. The accuracy of fraud detection is also high (79‒85 %). The results of our study confirm the effectiveness of using automatic machine learning algorithms to synthesize fraud detection models in digital payment systems. In this case, efficiency is manifested not only by the resulting classifiers' quality but also by the reduction in the cost of their development, as well as by the high potential of interpretability. Implementing the study results could enable financial institutions to reduce the financial and temporal costs of developing and updating active systems against payment fraud, as well as improve the effectiveness of monitoring financial transactions

Highlights

Data on the global financial statistics show that total losses from fraudulent transactions around the world are constantly growing
It should be taken into consideration that there is a trend around the world to introduce strategies to digitalize economic relations, in particular, the introduction of the “Bank-as-a-Service” concept by banks, which implies a significant increase in the burden on payment services
Our experiments have shown that in the context of operating a highly unbalanced sample of data, not all metrics were suitable for the procedure of automatic synthesis of machine learning models

Summary

Introduction

Data on the global financial statistics show that total losses from fraudulent transactions around the world are constantly growing. Hereinafter, we shall consider such transactions to be fraudulent that involve funds in the client’s account used by third parties, without the consent or permission of the account holder It should be taken into consideration that there is a trend around the world to introduce strategies to digitalize economic relations, in particular, the introduction of the “Bank-as-a-Service” concept by banks, which implies a significant increase in the burden on payment services. The digital nature of payment relationships has made it possible to significantly automate the development of models for identifying fraudulent transactions through the use of machine learning and Big Data analysis techniques, as shown by studies [3, 4], and others Even in this case, the development of a system to detect fraudulent transactions and its maintenance require considerable time, as well as attracting qualified professionals. Of additional interest is the assessment of the cost of model synthesis and training, as well as the analysis of the ability to interpret machine learning results

Literature review and problem statement

The aim and objectives of the study

Findings

Conclusions