Feature selection for chargeback fraud detection based on machine learning algorithms

Abstract

Fraud crimes continue to increase proportional to usage of electronic payments. Recently, many game users have been abusing the returns and refunds policy on Payment Gateway to obtain game money illegally. Although economic damage to game companies increases gradually, they cannot anticipate chargeback fraud for many reasons. Thus, it is necessary to create or select important features to detect anomalous users in advance. The goal of this paper is to predict abusers using machine learning techniques and the transaction data provided by a world-famous game company. We determine various features by preprocessing the given data. Comparative experiments are conducted using the WEKA tool, which is a collection of machine learning algorithms for data mining tasks. Several algorithms, such as decision tree and SVM, are used in the experiments. The experimental results show that SVM is usually superior to other methods. The results show over 30% improvement through SMOTE, an over-sampling technique. The major feature subsets, which were selected by using a simple genetic algorithm as a search method and Classifier Subset Eval as an attribute evaluator both based on DT classifier, consists of payment method processed in the latest, the count of transactions in the past a month, the standard deviation of charges, and the average of charges. © 2016, Research India Publications.