Search Network

Modelling Fraudulent Behaviour in E-Commerce

Executive summary

This project applied analytic models to identify fraudulent customers for ezetop, a company who provides an online mobile phone top-up service to over 130 mobile operators spanning 65 countries. Fraud occurs in their business when a fraudster uses a fake credit card to avoid paying for their service.

Challenge overview

Fraud is a major risk for ezetop and the objective of the project was to develop a prototype fraud detection system to help mitigate this risk.

Implementation of the initiative

ICHEC worked with ezetop staff to understand the fraud issues and gather the necessary information. A series of meetings helped to scope the problem and tailor the solution to ensure it would meet the company's needs. In seeking a solution it was essential to avoid classifying a good client as a fraudster and ensuring that any strategy used would minimise this particular classification error.

An initial cleaning of the data followed by an examination of the available parameters suggested a number of techniques that could best address the fraud issue. ezetop reviewed these suggestions and this led to the selection of a number of machine learning techniques. A key criterion was that the initial system should not be a black box but rather that the process would be visible and understandable to the company's fraud analysts. Over the course of the project this criterion was waived as the analysts became more familiar with the techniques.

We explored various techniques to tackle the fraudster problem and improve fraud detection within ezetop's existing business process. The project initially used a Decision Trees approach to provide visibility regarding the classification process. As the project progressed a Random Forest technique proved more suitable and this technique was more robust in minimising the risk of good clients being classified as fraudsters.

"We need to make a decision within 10 seconds because we need to deliver that credit to the mobile phone immediately ... ICHEC has provided us with a system where we can check every transaction that goes through .. ICHEC's system will be able to indentify whether a user is fraudulent or not, which is a huge benefit to our business," says David Bowles, Head of Online at Ezetop.

Fig.1. An example of a simplified Decision Tree for an online e-commerce site detecting fraudulent behaviour. (See Attachment)

Figure 1 shows how the decision making used in Decision Trees can be presented to analysts. This type of presentation helped improve their confidence with the techniques facilitating the move to the Random Forest technique later in the project. This technique uses an ensemble of Decision Trees with statistically controlled variation to facilitate model averaging and random feature selection. This forest of trees helps provides better accuracy and less risk that a good client is classified as a fraudster.

Results and achievements

By applying complex mathematical techniques we have delivered significant improvements to the company's fraud detection process. What this project demonstrated was that analytics can be a powerful tool for improving the company's existing business. As online ecommerce expands, fraud is becoming an active area of interest for those companies providing online services or products. The results of this work have encouraged other companies to use analytics to tackle similar problems in their business.

Lessons learned and replicability

The techniques used in this project can be applied to other aspects of e-commerce such as segment targeting or churn analysis. From the project's inception, we concentrated on finding the optimum learning algorithm to apply to the company's fraud problem.

 

Technology Transfer
Irish Centre for High End Computing (ICHEC)
http://www.ichec.ie/consultancy
consultancy@ichec.ie
David Bowles, Ezetop
http://www.ezetop.com
dbowles@ezetop.com

 

30 November 2011
print version share on facebook share on linkedin

Executive summary

This project applied analytic models to identify fraudulent customers for ezetop, a company who provides an online mobile phone top-up service to over 130 mobile operators spanning 65 countries. Fraud occurs in their business when a fraudster uses a fake credit card to avoid paying for their service.

Challenge overview

Fraud is a major risk for ezetop and the objective of the project was to develop a prototype fraud detection system to help mitigate this risk.

Implementation of the initiative

ICHEC worked with ezetop staff to understand the fraud issues and gather the necessary information. A series of meetings helped to scope the problem and tailor the solution to ensure it would meet the company's needs. In seeking a solution it was essential to avoid classifying a good client as a fraudster and ensuring that any strategy used would minimise this particular classification error.

An initial cleaning of the data followed by an examination of the available parameters suggested a number of techniques that could best address the fraud issue. ezetop reviewed these suggestions and this led to the selection of a number of machine learning techniques. A key criterion was that the initial system should not be a black box but rather that the process would be visible and understandable to the company's fraud analysts. Over the course of the project this criterion was waived as the analysts became more familiar with the techniques.

We explored various techniques to tackle the fraudster problem and improve fraud detection within ezetop's existing business process. The project initially used a Decision Trees approach to provide visibility regarding the classification process. As the project progressed a Random Forest technique proved more suitable and this technique was more robust in minimising the risk of good clients being classified as fraudsters.

"We need to make a decision within 10 seconds because we need to deliver that credit to the mobile phone immediately ... ICHEC has provided us with a system where we can check every transaction that goes through .. ICHEC's system will be able to indentify whether a user is fraudulent or not, which is a huge benefit to our business," says David Bowles, Head of Online at Ezetop.

Fig.1. An example of a simplified Decision Tree for an online e-commerce site detecting fraudulent behaviour. (See Attachment)

Figure 1 shows how the decision making used in Decision Trees can be presented to analysts. This type of presentation helped improve their confidence with the techniques facilitating the move to the Random Forest technique later in the project. This technique uses an ensemble of Decision Trees with statistically controlled variation to facilitate model averaging and random feature selection. This forest of trees helps provides better accuracy and less risk that a good client is classified as a fraudster.

Results and achievements

By applying complex mathematical techniques we have delivered significant improvements to the company's fraud detection process. What this project demonstrated was that analytics can be a powerful tool for improving the company's existing business. As online ecommerce expands, fraud is becoming an active area of interest for those companies providing online services or products. The results of this work have encouraged other companies to use analytics to tackle similar problems in their business.

Lessons learned and replicability

The techniques used in this project can be applied to other aspects of e-commerce such as segment targeting or churn analysis. From the project's inception, we concentrated on finding the optimum learning algorithm to apply to the company's fraud problem.

Technology Transfer
Irish Centre for High End Computing (ICHEC)
http://www.ichec.ie/consultancy
consultancy@ichec.ie
David Bowles, Ezetop
http://www.ezetop.com
dbowles@ezetop.com

 

Attached Documents

PDF icon Irish Mathematics and Industry (Irish_Mathematics_and_Industry.pdf | 1.42 MB)