CONTACT: GEORGE MCCRORY
100 Old Public Library
Iowa City IA 52242
(319) 384-0012; fax (319) 384-0024
e-mail: george-mccrory@uiowa.edu

Release: Aug. 30, 2000

UI management sciences team wins data mining competition

IOWA CITY, Iowa -- What kinds of people are likely to buy insurance policies for their trailers? The answer, submitted by assistant professor Nick Street and Ph.D. candidate YongSeog Kim of the University of Iowa Henry B. Tippie College of Business, helped the Iowa researchers win the CoIL Challenge, an international data mining competition.

Street and Kim of the college's department of management sciences won the description category of the CoIL Challenge, which was created to promote the application of data mining technology to real world problems. Data mining is a process of identifying valid and potentially useful patterns of information drawn from a large volume of data and is being used by companies to make informed decisions that affect the bottom line.

The competition consisted of two tasks derived from a Dutch insurance company's direct marketing data. One was to predict which customers would be interested in the policy, based on 86 variables including product usage data and socio-demographic data derived from postal codes. The other was to describe why customers have such a policy and how these customers are different from other people.

For the prediction task, Street and Kim used artificial neural networks (a machine learning approach inspired by the structure of neurons in the brain) and an evolutionary search method developed by Filippo Menczer, a UI assistant professor of management sciences, to choose a subset of the 86 variables that could reliably predict policy buyers. Their resulting model raised the expected hit rate of prospective policy buyers from under six percent to over 14 percent, a significant improvement that ranked with the top handful of entries in the category.

For the description task, Street and Kim combined information from an artificial neural network with other data mining techniques to build a descriptive profile of likely -- and unlikely -- customers. These models learn to spot relationships in the data that may escape human inspection.

So who are the most likely customers to buy insurance for trailers? They come from upper-middle-class families with more than one insured automobile, and possibly other types of insurance policies. Bad bets include both extremes of the income scale, farmers, and somewhat surprisingly, senior citizens.

Their entry was chosen as the winner from 41 entries submitted from around the world by both academic and professional contestants, as judged on comprehensibility, usefulness and actionability.

According to Stephan van Heusden, a marketing consultant from MSP Associates who judged the contest, a good entry combined a description of the results with a 'tool' to use the results. He said the UI team "clearly explained" which steps preceded conclusions. "(They) may have discovered all conclusions the others also found," but additionally, tried to interpret them, said van Heusden.

For more information, contact Street at (319) 335-1016 or nick-street@uiowa.edu, and see the CoIL challenge website at http://www.dcs.napier.ac.uk/coil/.