Behavior: The Next Marquee ApplicationIt's time to start using more advanced analyticsContinued from Page 1 At the same time, our data sources have descended to the subtransactional level. It seemed in the late '90s that if we captured every atomic sales transaction, we had somehow arrived at a fundamental foundation for all possible data. But the development of CRM and the capture of presales customer behavior has opened up potentially another factor of 1,000 in the amount of data we can collect about our customers. These new data sources include individual page requests tracking visitors on the Web, call center logs relating to product information and customer support, market basket information from retail and financial companies, and promotion response tracking. We are just in the beginning of explosive growth of these subtransactional data sources. Soon we will have global positioning systems embedded in our cars, our passports, and our credit cards. At the same time, our increased security needs will allow us to see customers coming and going from many of our stores and offices. I'll sidestep the legitimate issues of privacy raised by these technologies. The marquee data warehousing application in the 2000s, in my opinion, will be customer behavior. We will analyze both individual and commercial customer behavior. But what is behavior, exactly? It certainly doesn't seem to be as simple as shipments, share, or profit. What does it mean to add up behavior? Is behavior even numeric? The New Analytics of BehaviorAlthough Michael Berry and Gordon Linoff haven't used the word "behavior" in the titles of their books, it's easy to cast what they've written as a foundation for the large topic of behavior. In their latest book, Mastering Data Mining, The Art and Science of Customer Relationship Management (Wiley, 2000) they show how the simple progression of clustering, classification, and prediction takes subtransactional data sources and turns them into actionable descriptions of behavior. Briefly, clustering is the recognition of discrete cohort groups (typically of customers) from the ocean of all customers. Clustering can be accomplished with a number of different advanced data mining techniques described by Berry and Linhoff. Note the cultural jump required here: You have to trust the clustering algorithms. Classification is possible once you have clusters. If a new customer prospect can be associated with one of your existing clusters, you can reasonably infer that this customer will behave like the other members of that cohort group. Note the word behave. You have classified the customers by their behavior. We need some more advanced analytics here in order to understand how close the prospect is to the centroid of the existing cohort group, another step up the analytic ladder. Prediction is the highest art form. You can associate a numeric metric with each known member of a cohort group and then use that metric together with the "distance" to the new prospect to derive a numeric prediction of lifetime value, or likelihood to default. At any given point in time, the behavior of a custoomer can be summarized by a textual tag, such as Regular High Margin Customer, or maybe Unproductive Window Shopper. One of the chief goals of the marketing manager is how to convert customers from this second group into the first group. Lately I have been helping a number of data warehouse designers create schemas for tracking and reporting exactly this kind of label transition over time in their customer-oriented data warehouses. Although the final reports are pretty straightforward, I hope you can appreciate the significant analytic foundation they require. Finally, a tricky but very compelling form of behavior is understanding the paths taken by customers as they visit your Web site or otherwise access all the customer communication points of your organization. Data miners call this link analysis. Again, once you diagnose the path's significance, you can assign it a behavioral label, and then use the techniques I just described to boil it down to an actionable report for the marketing manager. In my next column, I'll write about the challenges of converting free-form text data into behavior tags that can then participate in a normal data warehouse. But meanwhile, go read Berry and Linhoff's book. Ralph Kimball co-invented the Star Workstation at Xerox and founded Red Brick Systems. He has three best-selling data warehousing books in print, including The Data Webhouse Toolkit (Wiley, 2000). He teaches dimensional data warehouse design through Kimball University and critically reviews large data warehouse projects. You can reach him through his Web site, www.rkimball.com.
|
 |
Most Popular This Week
IE Weekly Newsletter
Subscribe to the newsletter
| |
| ||||||||||||||||||||||||||||||||
