2013 credit card fraud detection why theory dosent adjust to practice

Copyright © 2013, SAS Institute Inc. All rights reserved. #analytics2013
Credit Card Fraud Detection
Why Theory Doesn't Adjust to Practice
Alejandro Correa Bahnsen, Luxembourg University
Andrés Gonzalez Montoya, Scotia Bank

Introduction
€ 500
€ 600
€ 700
€ 800
2007 2008 2009 2010 2011E 2012E
Europe fraud evolution
Internet transactions (millions of euros)

Introduction
$-
$1.0
$2.0
$3.0
$4.0
$5.0
2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
US fraud evolution
Online revenue lost due to fraud (Billions of dollars)

• Increasing fraud levels around the world
• Different technologies and legal requirements makes
it harder to control
• There is a need for advanced fraud detection
systems
Introduction

• Introduction
• Transaction flow
• Database
• Evaluation of algorithms
• If-Then rules (Expert Rules)
• Financial measure
• Predictive modeling
• Logistic Regression
• Cost Sensitive Logistic Regression
Agenda

Simplify transaction flow
Fraud??
Network

Data
• Larger European card
processing company
• 2012 card present transactions
• 750,000 Transactions
• 3500 Frauds
• 0.467% Fraud rate
• 148,562 EUR lost due to fraud
on test dataset
Dec
Nov
Oct
Sep
Aug
Jul
Jun
May
Apr
Mar
Feb
Jan
Test
Train

• Raw attributes
• Other attributes:
Age, country of residence, postal code, type of card
Data
TRXID Client ID Date Amount Location Type
Merchant
Group
Fraud
1 1 2/1/12 6:00 580 Ger Internet Airlines No
2 1 2/1/12 6:15 120 Eng Present Car Rent No
3 2 2/1/12 8:20 12 Bel Present Hotel Yes
4 1 3/1/12 4:15 60 Esp ATM ATM No
5 2 3/1/12 9:18 8 Fra Present Retail No
6 1 3/1/12 9:55 1210 Ita Internet Airlines Yes

• Derived attributes
Data
Trx
ID
Client
ID
Date Amount Location Type
Merchant
Group
Fraud
No. of Trx – same
client – last 6 hour
Sum – same client
– last 7 days
1 1 2/1/12 6:00 580 Ger Internet Airlines No 0 0
2 1 2/1/12 6:15 120 Eng Present Car Renting No 1 580
3 2 2/1/12 8:20 12 Bel Present Hotel Yes 0 0
4 1 3/1/12 4:15 60 Esp ATM ATM No 0 700
5 2 3/1/12 9:18 8 Fra Present Retail No 0 12
6 1 3/1/12 9:55 1210 Ita Internet Airlines Yes 1 760
By Group Last Function
Client None hour Count
Credit Card Transaction Type day Sum(Amount)
Merchant week Avg(Amount)
Merchant Category month
Merchant Country 3 months
– Combination of following criteria:

• Misclassification = 1 −
𝑇𝑃+𝑇𝑁
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
• Recall =
𝑇𝑃
𝑇𝑃+𝐹𝑁
• Precision =
𝑇𝑃
𝑇𝑃+𝐹𝑃
• F-Score = 2
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∗ 𝑅𝑒𝑐𝑎𝑙𝑙
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙
Evaluation
True Class (𝑦𝑖)
Fraud (𝑦𝑖=1) Legitimate (𝑦𝑖=0)
Predicted class
(𝑝𝑖)
Fraud (𝑝𝑖=1) TP FP
Legitimate (𝑝𝑖=0) FN TN
• Confusion matrix

Fraud
Algorithms
• If-Then rules
• Predictive modeling
• Logistic Regression
• Decision Trees
• Random Forest
• Cost Sensitive
Logistic Regression
Fraud??
Network

• “Purpose is to use facts and rules, taken from the
knowledge of many human experts, to help make
decisions.”
• Example of rules
• More than 4 ATM transactions in one hour?
• More than 2 transactions in 5 minutes?
• Magnetic stripe transaction then internet transaction?
If-Then rules (Expert rules)

• More than 4 ATM transactions in one hour?
• More than 2 transactions in 5 minutes?
• Magnetic stripe transaction then internet
transaction?
Fraud??
Network
If one or more rules is activated then decline the transaction

• Problems with rules
• New fraud patterns are not detected
• Only simple rules can be created
• Advantages of rules
• Easy to implement
• Very easy to interpret

1.04%
31%
17%
22%
Miss-cla Recall Precision F1-Score
Results

• Motivation
• False positives carries a different cost than
false negatives
• Frauds range from few to thousands of euros
(dollars, pounds, etc)
Financial evaluation
There is a need for a real comparison measure

• Cost matrix
where:
• Evaluation measure
Financial evaluation
Ca Administrative costs
Amt Amount of transaction i
Predicted class
(𝑝𝑖)
Fraud (𝑝𝑖=1) Ca Ca
Legitimate (𝑝𝑖=0) Amt 0

If-Then rules
1.04%
31%
17%
22%
Results
€
95,520
€
148,562
Cost Cost No Model
148,562 EUR are the losses due to fraud in the test database (2 months)

Predictive modeling is the use of statistical and
mathematical techniques to discover patterns in data in
order to make predictions
Predictive modeling

Predictive modeling
Amountoftransaction
Number of transactions last day
Normal Transaction
Fraud

Predictive modeling
Amount of transaction
Number of transactions last day
Normal Transaction
Fraud
Amount spend on internet last month

Predicted class
(𝑝𝑖)
Fraud (𝑝𝑖=1) 0 1
Legitimate (𝑝𝑖=0) 1 0
• Model
• Cost Function
• Cost Matrix
Logistic Regression

€
148,196
€
148,562
Cost Cost No Model
0.52% 0%
2%
0%
Logistic Regression
Results
148,562 EUR are the losses due to fraud in the test database (2 months)

1% 5% 10% 20% 50%
Logistic Regression
Sub-sampling procedure:
0.467%
Select all the frauds and a random sample of the legitimate transactions.
620,000
310,000
62,000
31,000 15,500 5,200
Fraud Percentage

Logistic Regression
Results
€ 148,562 € 148,196
€ 142,510
€ 112,103
€ 79,838
€ 65,870
€ 46,530
€ -
€ 20,000
€ 40,000
€ 60,000
€ 80,000
€ 100,000
€ 120,000
€ 140,000
€ 160,000
0%
10%
20%
30%
40%
50%
60%
70%
No Model All 1% 5% 10% 20% 50%
Cost Recall Precision Miss-cla F1-Score
Selecting the algorithm by Cost

Logistic Regression
• Best model selected using traditional F1-Score does not gives the best results in
terms of cost
• Model selected by cost, is trained using less than 1% of the database, meaning there
is a lot of information excluded
• The algorithm is trained to minimize the miss-classification (approx.) but then is
evaluated based on cost
• Why not train the algorithm to minimize the cost instead?

Predicted class
(𝑝𝑖)
Fraud (𝑝𝑖=1) Ca Ca
Legitimate (𝑝𝑖=0) Amt 0
• Cost Matrix
Cost Sensitive Logistic Regression
• Cost Function
• Objective
Find 𝜃 that minimized the cost function (Genetic Algorithms)

• Cost Function
• Gradient
• Hessian

0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Legitimate
Fraud
Amount cumulative distribution
€49
€370€124
€196

€ 148,562
€ 31,174
€ 37,785
€ 66,245 € 67,264
€ 73,772
€ 85,724
€ -
€ 20,000
€ 40,000
€ 60,000
€ 80,000
€ 100,000
€ 120,000
€ 140,000
€ 160,000
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
No Model All 1% 5% 10% 20% 50%
Cost Recall Precision F1-Score
Cost sensitive Logistic Regression
Results

Cost sensitive Logistic Regression
Results€ 148,562
€ 95,520
€ 46,530
€ 31,174
€ 35,466 € 34,203
€ -
€ 20,000
€ 40,000
€ 60,000
€ 80,000
€ 100,000
€ 120,000
€ 140,000
€ 160,000
0%
10%
20%
30%
40%
50%
60%
70%
80%
No Model If-Then rules Logistic Regression Cost Sensitive
Logistic Regression
Decision Trees Random Forests
Cost Recall Precision F1-Score

Conclusion
• Selecting models based on traditional statistics does not
gives the best results in terms of cost
• Models should be evaluated taking into account real
financial costs of the application
• Algorithms should be developed to incorporate those
financial costs

Contact information
Alejandro Correa Bahnsen
University of Luxembourg
Luxembourg
al.bahnsen@gmail.com
http://www.linkedin.com/in/albahnsen
http://www.slideshare.net/albahnsen

Thank You!!
Alejandro Correa Bahnsen
Andres Gonzalez Montoya

• Hastie, T., & Tibshirani, R. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction.
Beijing.
• Hand, D., Whitrow, C., Adams, N. M., Juszczak, P., & Weston, D. (2007). Performance criteria for plastic card fraud
detection tools. Journal of the Operational Research Society, 59, 956–962.
• Sheng, V., & Ling, C. (2006). Thresholding for making classifiers cost-sensitive. Proceedings of the National
Conference on Artificial Intelligence.
• Bhattacharyya, S., Jha, S., Tharakunnel, K., & Westland, J. C. (2011). Data mining for credit card fraud: A
comparative study. Decision Support Systems, 50(3), 602–613.
• Ling, C., & Sheng, V. (2008). Cost-sensitive learning and the class imbalance problem. In C. Sammut & G. I. Webb
(Eds.), Encyclopedia of Machine Learning (pp. 231–235). Springer.
• Moro, S., Laureano, R., & Cortez, P. (2011). Using data mining for bank direct marketing: An application of the
crisp-dm methodology. In EUROSIS (Ed.), European Simulation and Modeling Conference - ESM’2011 (pp. 117–
121). Guimares, Portugal.
References

2013 credit card fraud detection why theory dosent adjust to practice

Recomendados

Recomendados

Más contenido relacionado

La actualidad más candente

La actualidad más candente (20)

Destacado

Destacado (10)

Similar a 2013 credit card fraud detection why theory dosent adjust to practice

Similar a 2013 credit card fraud detection why theory dosent adjust to practice (20)

Más de Alejandro Correa Bahnsen, PhD

Más de Alejandro Correa Bahnsen, PhD (6)

2013 credit card fraud detection why theory dosent adjust to practice