KDD-2001 TutorialValue-based Data Mining for CRM
|
|||
| Abstract | Presenters | ||
|
One of the reasons that Customer Relationship Management (CRM) has not achieved its full potential is lack
of focus upon economic value and profitability. This tutorial surveys 3 important CRM data mining tasks:
attrition modeling, cross-sell, and web mining, and shows how these tasks should be modified to increase ROI.
To take one example, a customer retention campaign might execute extremely well from a technical point of view, but if the customers who are being saved are low-profit, then the overall result of the campaign might be negative. Another example is "partial" attrition: if a big customer withdraws a large part of their account, this is more important than when a small customer closes their account. For this reason, a value attrition approach is usually more profitable than traditional account attrition. Initial results suggest that value attrition models can capture 5-10 times more value than traditional account attrition models. For building cross-sell models, one needs to consider proper product definitions, business rules that define allowed product/customer interactions, modeling propensity to buy, estimating product profitability, and finally optimization for meeting enterprise goals. We discuss critical issues including converting model scores into probabilities and avoiding self-selection bias, and present 7 steps to global cross-sell optimization. With web mining, the key is to discover meaningful web behavior that can improve the financial performance of an e-business. E-Metrics are the metrics for measuring web site behavior and the indicators of e-business effectiveness. In addition to core web measurements such as hits, page views, visits, and users, and traditional measurements such as recency, frequency, and monetary value, E-Metrics are extended to stickiness, focus, migration rate and freshness. Framing proper web mining tasks with respect to E-Metrics and reporting E-Metrics with respect to web mining results are critical to true success of web mining. Other key issues include making web mining results understandable and actionable by business users, and integrating web behavior with offline customer data. |
Steve Gallant, Ph.D. - Director, Analytic Services, Xchange Inc. Leads predictive analytic consulting engagements for
Xchange. Formerly, Senior Scientist at HNC Software and Belmont Research, and Associate Professor of Computer Science at Northeastern University. Author of Neural Network Learning (MIT Press); three patents in neural networks, expert systems and information retrieval; and 40 published papers. |
||