November 21 , 2002

Like nervous students who know the right answer but aren't sure why, computers sometimes know more than they "think."

For example, a computer using a "neural network" can often reliably pick high performing stocks. But, like the student asked to explain his answer, the computer often can't articulate how it came up with the answer. It just "knows."

"But if you want people to make a decision based on the output, they need to know 'why,'" said management information science professor Monica Lam.

In search of "why," Lam developed an innovative way to coax the answer from the computer.

"Now we can explain to customers why we are making a recommendation," she said. And, she added, that explanation often leads to even more accurate predictions.
While neural networks aren't new-the first mathematical models were developed in the 1950s-they have become more sophisticated and are being used in a wider range of applications. The networks are modeled after the processes that allow mammals (like that tongue-tied student) to think and learn.

Like people, neural networks learn through training and can apply learning to new situations. The networks are first given data with a known outcome.

In the stock-picking example, company financial data would be input along with the company's stock performance. The network then processes the data to identify the "rules" that connect the financial data to stock performance. In this example, the network is taught to recognize high, average and low performing stocks. Once it has established the rules, the neural network can analyze any stock and predict stock performance.

Stock selection is a common use of neural networks. Yet investors are often reluctant to invest based on the abstract calculations in a black box, especially since neural networks have not been able to explain the rules.

"The traditional neural networks couldn't do that for you," Lam said. Now they can. Lam created an algorithm that queries the neural networks and reveals the rules the system developed to evaluate the data and make predictions.

"It will figure out what relationships are established during the learning process," she said. "Then I can judge whether I want to use the results."

Even better, because her algorithm looks for the most significant relationships, it filters out extraneous or insignificant information.

"The rules then have a higher prediction ability than the neural network itself," Lam said.

While Lam looked specifically at stock picking, neural networks and the algorithm could be applied to a variety of problems in business and other fields-including medicine.

Her work was published in the IEEE Transactions on Knowledge and Data Engineering and the proceedings of the first International Conference on Electronic Business last year. She holds a copyright on the computer code that implements her algorithm, though the algorithm itself is in the public domain.

####