Machine-Learning Methods for Credit Card Fraud Detection

Abstract

In order to thwart fraudsters, financial institutions must use current, advanced, customized predictive analytics to protect themselves. Data scientists and statisticians who understand machine learning and statistical methods are in increasingly high-demand and the demand for them is growing each year. Technically, machine learning is a subfield of artificial intelligence whereas statistics is subdivision of mathematics and many believe they only need in depth knowledge of one in order to be a predictive modeler. This fallacy leads to inefficient and/or inaccurate models, and sadly, many industries have not yet realized that the mathematics behind the model is just as important, if not more important, than the computer science needed to implement it. However, some businesses have and this thesis will hopefully help both industry and academia move further along in this direction.

In this thesis, we explore existing methodologies for fraud detection proposed by academic professionals around the globe and illustrate their accuracy, efficiency and reliability on a large dataset downloaded from a public website. The methods analyzed are hidden Markov models (HMM), convolutional neural networks (CNN), and support vector machines (SVM). For each method, we present the history and motivation, theoretical framework, strengths and weaknesses, and numerical examples done in either R or SAS Enterprise Miner.