Stimulation Initiative for European Neural Applications
Esprit Project 9811
Fraud detection in credit card transactions
Company background
SEMP is the company that deals with all the activities related to and
originated by VISA credit cards issued by Spanish banks. A typical value of
the total amount of such transactions is between 500.000 and 1.000.000
operations per day.
Description of the problem
The fight against fraud in transactions with VISA cards is obviously a
problem of highest interest for SEMP. The main sources of fraud are theft
and forgery of credit cards, and transactions with a higher economical
impact must be processed on-line in order to obtain clearance. Systems with
on-line capacity to detect if a transaction is fraudulent are therefore a
prioritary objective for SEMP. The complexity of the inherent operation
characteristics make very problematic the viability of rule-based
applications. Besides, the changing nature of fraud would make specially
difficult and expensive the maintenance of such a system.
Neural Networks techniques application
Distinction between a legal operation and a fraudulent one can be seen as a
problem of two-category classification. Given the known ability of neural
nets to carry out non linear classifications among sets, they have a natural
application to the problem under consideration. The examples used in the
construction of the detection model were taken from the analysis of the year
1994 and the first half of 1995 traffic on the operating sectors of main
fraud impact (with a total of more than 3 million transactions). The net,
developed in the "Instituto de Ingenieria del Conocimiento" (Knowledge
Engineering Institute), implements what essentially is Fisher's
discriminant
analysis, after a non-linear transform of the detection variables.
Benefits
The natural measure of the system success is obviously its ability to
decrease fraud, expressed as the resulting economical savings. The system
assumes a decrease of fraud of about 30-40% on the main impact sectors.
Generalization
The detection problem considered here is characteristic of a wide family of
decision problems for which the rule-based systems have difficult
application and require expensive maintenance, while neural systems like the
one described provide a fast construction of models and a relatively easy
adaptation of these models to pattern changes. The applicability range of
the techniques used in the SEMP problem is very wide, with particular
interest in sectors like banking or insurance.
Contact person
Francisco Ginel, SEMP, c/. Gustavo Fernandez Balbuena 15, 28002 Madrid
References
C. Santa Cruz et al., "Hybrid neural methods in classification problems",
IIC Technical Report 9501, Instituto de Ingenieria del Conocimiento,
Universidad Autonoma, 28049 Madrid.