In-Depth
Mellon Banks on Mining
Increasing revenues, reducing risks and maintaining a competitive edge require sophisticated tools and techniques. Pittsburgh-based Mellon Bank Corporation had been using data mining for its marketing activities for years and has recently expanded its use to fraud prevention, delinquency prediction and business process reengineering, including streamlining loan underwriting. But its most significant use of data mining, by far, is managing its complex customer relationships.
"Our overall strategy is to provide superior customer service," says Peter Johnson, Mellon’s Vice President for the Advanced Technology Group. "And one way we do that is through data mining. What we’re trying to do as part of data warehousing is conduct target and one-to-one marketing."
"Data mining is the carrot that justifies the expensive stick of data warehousing, and data warehousing is the enabler for data mining," explains Johnson. "Generally, an organization spends 80 percent of its time culling data from various sources. A data warehouse can streamline the data mining process. For example, if I want to do a predictive model for credit card attrition, I can do it in a few days instead of a few months by using data warehousing." Based on the data mining results, the bank can take action to retain customers’ loyalty.
Johnson believes that data mining is the highest level of analysis. "It is very proactive, but you also find things you didn’t expect that help your competitive advantage," says Johnson.
You can perform data mining without data warehousing. "We were doing data mining before data warehousing was invented," says Johnson, "but without the warehouse, data mining becomes an ad hoc special project requiring a fair amount of programming and other efforts to pull data from various sources." In the last few years, Mellon has invested significantly in data warehousing, building its enterprise warehouse in DB2 on OS/390. It plans to make data mining a normal part of its operation within the next few years. The bank uses Business Objects on an NT Web Server to access the DB2 data warehouse on the S/390.
Although several vendors provide attrition models, the bank decided it was an important part of its business strategy and wanted the capability to build its own models. Therefore, in 1995, it began a joint project with IBM, focusing on a broad-based multi-platform data mining tool, called Intelligent Miner for Data. Intelligent Miner for Data, Version 2.1 enables users to mine structured data stored in conventional databases or flat files. The software tool searches for hidden information stored in databases, data warehouses and data marts.
"We deploy data mining tools on an ongoing basis," says Johnson. "The organization already had the basic data mining units, so when we gave them another tool, they could do their jobs better and more efficiently. As with any given tool, users have to learn how to use it, but adopting Intelligent Miner required no major shift in job responsibility."
Recently the bank has been putting Intelligent Miner for Data, Version 2 to the test. "Data Miner scales very well against warehoused data," says Johnson. "That’s the biggest reason for implementing it. IBM’s data mining tool was designed to work against data in a database without having to be in a special format, so the biggest advantage for us is its integration with our warehousing strategy. When we work in SAS, for example, we have to jump through hoops."
That’s only one reason why Intelligent Miner is so well received. "The technology is relatively mature," states Johnson. "The graphical user interface is intuitive and easy to use within both Unix and Windows clients. This will enable us to deploy the tool to a wider range of users throughout our enterprise." Previously, analysts at Mellon had to rely on its Information Technology specialists to perform data mining runs for them. Now they can do it themselves from their Windows NT desktops.
In optimizing its statistical functions, Intelligent Miner expedites the data mining process. The principal component analysis function summarizes data and helps reduce the number of variables required for various analyses.
Mellon Bank was familiar with the technology of data mining before IBM had a product. The bank tested Intelligent Miner with large quantities of data at IBM’s Teraplex in Poughkeepsie, N.Y., a laboratory for customers to use to understand the technology in an environment that reduces risk on both sides. "Data mining is incredibly computational," maintains Johnson. "We were able to build a predictive model of customer behavior based on one million rows and more than 600 columns of data. Where it performed better than expected was scalability. Intelligent Miner allows us to analyze larger datasets than possible before," states Johnson.
But according to Johnson the technology falls short, in certain respects. "When you use data mining tools the software expects the data to be in one table," says Johnson. "If I want to predict customer behavior, everything I have to know about the customer has to be represented in a single table, but our data structure is much more complicated. In reality, the data consists of numbers and tables with relationships. If you’re a Mellon customer and you have multiple products we know all your product relationships, what household you’re in, demographics, as well as credit card transactions. Therefore, we have to do pre-processing, which is part of the art and technique of data mining."
Johnson concludes, Data mining specialists have to intervene to take complex data relationships and represent that history of customer behavior in simpler terms. The tool can take over from there, but it can’t automate that process."