Credit Card Fraud Classification

1
Credit Card Fraud Classification
Modern Approaches and Practices

• May 4, 2005

ENEE 752 Spring 2005
Gavin Rosenbush
2
Credit Card Fraud Classification
Introduction

• The credit industry is extremely large and
  ubiquitous
• Almost everyone has at least one line of credit
• Credit institutions handle millions of
  transactions per day
• Each transaction is checked by the credit
  institutions and
• is denied or approved

ENEE 752 Spring 2005
Gavin Rosenbush
3
Credit Card Fraud Classification
Introduction

• Why do we care about fraud?

• Fraud is paid for by consumers
• Banks raise rates to vendors and consumers
• Vendors raise prices
• In the end, buyers take responsibility

ENEE 752 Spring 2005
Gavin Rosenbush
4
Credit Card Fraud Classification
Domain Specification

• What is credit card fraud?

• Generic def'n unauthorized use of someone's
  credit card for
• making purchases
• There are many cases/situations that are
  fraudulent
• Well-known person's physical card stolen
• Information unknowingly sniffed
• Social Engineering
• Corporate databases compromised
• Many points of failure in the system
• Physical card not required for fraud

ENEE 752 Spring 2005
Gavin Rosenbush
5
Credit Card Fraud Classification
Domain Specification

• What is credit card fraud?

• Because of all these cases, detecting fraud when
• information stolen is difficult.
• Fraud must be detected when transactions occur
• Combined with the volume of daily transactions,
  machine
• Learning becomes necessary for this
  classification

ENEE 752 Spring 2005
Gavin Rosenbush
6
Credit Card Fraud Classification
System Description

• Information to be deduced

• Two class classification fraudulent, legitimate
• System should output information to assist in
  this decision
• Different algorithms have different output
  conventions
• Probability that transaction is fraudulent
• Probability that transaction is legitimate

ENEE 752 Spring 2005
Gavin Rosenbush
7
Credit Card Fraud Classification
System Description

• What cannot be deduced

• Machine learning algorithms are not able to
  provide a
• reason why a certain decision was made
• Not required in many other problems, but would be
  useful in
• this problem. Customers will ask why actions
  occurred
• A person with domain expertise may be able to
  look at
• the case and provide a reason

ENEE 752 Spring 2005
Gavin Rosenbush
8
Credit Card Fraud Classification
System Constraints

• What makes this problem unique?

• Skewed class distribution
• Difference in percentage of fraudulent and
  legitimate
• transactions is very high
• Non-Uniform Cost Function
• The cost of incorrectly or correctly classifying
  a transaction
• is not uniform and is based on cost of the
  transaction and
• the customer's history

ENEE 752 Spring 2005
Gavin Rosenbush
9
Credit Card Fraud Classification
System Constraints

• What makes this problem unique?

• Overlapping data
• Legitimate transactions can appear fraudulent
  vice-versa
• Speed demand is high
• Customers don't want to wait more than a few
  seconds
• False-Positive rate
• FP rate is very important customers don't want a
  hassle
• Data sharing is not encouraged
• Customer privacy
• Trade secrets

ENEE 752 Spring 2005
Gavin Rosenbush
10
Credit Card Fraud Classification
Research Constraints

• What made this difficult to research?

• Little information on in-use systems
• Fortunately, one was studied
• Training data kept secret
• Features kept secret

ENEE 752 Spring 2005
Gavin Rosenbush
11
Credit Card Fraud Classification
System Analysis

• Introduction

• Researched the following
• Karl Tuyls' comparison of neural networks and
  bayesian
• networks
• Minerva
• Spain's neural-network based system for
  classifying VISA
• transactions

ENEE 752 Spring 2005
Gavin Rosenbush
12
Credit Card Fraud Classification
System Analysis

• Common Elements

• Two main system types
• By-owner system
• History of the card owner stored and compared
  against
• current transaction
• Large data storage requirement
• By-operation system
• History of operations stored and compared over a
  fixed
• time window
• Smaller storage requirement, faster processing

ENEE 752 Spring 2005
Gavin Rosenbush
13
Credit Card Fraud Classification
System Analysis

• Common Elements

• Input Features
• Not specified due to privacy
• Hypothesis looked at paper on credit
  application fraud
• Features are characteristics about the current
  transaction
• Most likely features (among others)
• Cost of transaction
• Time of day
• Location
• Store name, type

ENEE 752 Spring 2005
Gavin Rosenbush
14
Credit Card Fraud Classification
Artificial Neural Networks

• Applied to credit fraud

• Most widely used method
• In use system Minerva
• Handles 60 of VISA traffic in Spain
• 75 of Spain's credit card institutions
• More than 1.2 million transactions per day
• By-operation system

ENEE 752 Spring 2005
Gavin Rosenbush
15
Credit Card Fraud Classification
Artificial Neural Networks

• Applied to credit fraud

• Network Structure
• 1 input layer
• 1-2 hidden layers 3-6 nodes each
• 1 output layer
• Activation Functions
• Sigmoid most widely used
• Hyperbolic tangent function also used with
  comparable
• results

ENEE 752 Spring 2005
Gavin Rosenbush
16
Credit Card Fraud Classification
Artificial Neural Networks

• Applied to credit fraud

• Pre-Processing Data Improves ANN performance
• Restricting input features to those that are most
  relevant
• Requires domain expertise
• Normalization of input features
• De-correlation of input features to eliminate
  unnecessary
• Features
• ANNs with preprocessing represented the best
  results

ENEE 752 Spring 2005
Gavin Rosenbush
17
Credit Card Fraud Classification
Bayesian Belief Networks

• Applied to credit fraud

• Comparable results
• Several difficulties in comparison
• Less features used
• Different training, test data used
• Different network structures tried
• Evaluated using STAGE algorithm

ENEE 752 Spring 2005
Gavin Rosenbush
18
Credit Card Fraud Classification
Data Sharing using Meta-Learning

• Applied to credit fraud

• JAM Java Agents for meta-learning
• Distributed computing model at the OS level
• Allows each institution to have their own local
  classifier
• Java agents work with local classifiers at each
  site to
• share learned information
• Secured environmentallows learning info to be
  shared
• Without divulging personal or proprietary info

ENEE 752 Spring 2005
Gavin Rosenbush
19
Credit Card Fraud Classification
System Comparisons

• System constraints are key in comparisons
• True positive rate
• False positive rate
• Speedboth learning and classifying new examples
• Storage space requirements

ENEE 752 Spring 2005
Gavin Rosenbush
20
Credit Card Fraud Classification
Results

• ANNs

• Best results with pre-processing
• Training 120K transactions, Test 100K
  transactions
• Overfitting occurs after 90 epochs
• At /- 10 false positive, 60 true positive
• At /- 15 false positive, 70 true positive
• Minerva
• 111 false to true positive rate at best
• 12 false to true positive rate at worst

ENEE 752 Spring 2005
Gavin Rosenbush
21
Credit Card Fraud Classification
Results

• ANNs

• Execution time-
• Fast
• Minerva reported at 60ms, dominated by
• disk access
• Training time -
• Slow about 2 hours

ENEE 752 Spring 2005
Gavin Rosenbush
22
Credit Card Fraud Classification
Results

• BBNs

• Comparable performance ratios to ANNs
• Execution time-
• Slow
• Training time -
• Fast about 20 minutes

ENEE 752 Spring 2005
Gavin Rosenbush
23
Credit Card Fraud Classification
Conclusion

• ANNs win

• Neural networks are more widely used because
  speed is
• critical and performance is comparable

ENEE 752 Spring 2005
Gavin Rosenbush
24
Credit Card Fraud Classification
Future Work

• There is a lot of work still to be done in the
  field
• Information sharing (JAM) should assist greatly
• Pruning algorithms for removing unnecessary
  perceptrons
• Support Vector Machines have not been considered

ENEE 752 Spring 2005
Gavin Rosenbush
25

• Questions
• Possible Answers