Mellon Bank has bought software that uses an advanced form of artificial intelligence to detect credit card fraud, company officials said.
The software, called Fraud Detection System, uses a type of artificial intelligence programming called neural networking, is marketed by Nestor Inc., Providence, R.I.
A number of banks have been developing neural network systems to help detect fraud and manage risk in consumer lending. Adherents of the technology claim neural networking can radically improve accuracy and productivity.
Chase Manhattan Corp., Citicorp, and Security Pacific Corp. all have begun neural network projects over the past two years. But projects at Citicorp and Security Pacific are currently on hold.
Brain Function Simulated
Neural networking is a branch of artificial intelligence that is designed to discern patterns in large amounts of data in a way that simulates the neural connections of the human brain. The software can change the weights it assigns to variables automatically, as if learning from experience.
The product from Nestor is the first commercially available neural network for fraud detection. One other customer, a European bank whose name the vendor would not disclose, already has the software in place.
Although Mellon officials declined to say how much the bank paid for the software, it is said to cost $30,000 to $50,000 for a small-scale pilot model and more than $300,000 for a full-blown system.
Mellon expects to test and possibly deploy the software in the spring.
Credit Authorization Role
If the test is successful, Mellon will install the software immediately in the credit authorization area. Mellon has a $1.1 billion credit cart portfolio.
The neural network technology will help Mellon boost productivity and improve its defenses against credit card fraud, said Philip Samson, vice president and portfolio performance manager at Mellon. Credit card fraud at the company climbed 50% in 1991 over 1990, Mr. Samson said.
Mellon is now modifying the system to reflect the types of behavior it wants to follow. The software will be programmed to detect changes in behavior, such as frequency of credit card use, types of purchases, and size of payments.
The system will look for patterns indicating specific scams, like a common one in which a cardholder makes large payments to inflate a credit line just before borrowing to the limit and absconding.
In developing the system, Mellon has reviewed six months' worth of credit card transactions, identifying fraudulent ones. It has merged that information with six months' worth of statement information to determine patterns of fraudulent transactions.
Neural networks can help a bank catch a higher number of fraud attempts, and with less work. Ken Liu, vice president of HNC Inc., a San Diego-based vendor of neural networks, said that under normal circumstances, a bank can catch about 10% of the credit card crimes perpetrated.
But in order to catch this small percentage of cheats, the bank must call about 400 customers. With one neural network system that HNC installed, the bank caught about 20% of the fraud, and made only 40 calls. Mr. Liu declined to name the bank.
Although neural network technology has been talked about for many years, it is still in its infancy. Neural networks have proved more complex to develop than banks and vendors realized.
"Right now there's a lot of black magic in terms of what paradigms are used," said Neena Buck, a consultant with New Science Associates, Southport, Conn., which follows the artificial intelligence market.