This is one of the most important questions to ask when deciding how best to translate a business problem into a data mining problem. Surprisingly often, the initial answer is “we’re not sure.” An answer is important because, as the cautionary tale in the sidebar illustrates, different intended uses dictate different solutions.
For example, many of our data mining engagements are designed to improve customer retention. The results of such a study could be used in any of the following ways:
■■ Proactively contact high risk/high value customers with an offer that rewards them for staying.
470643 c03.qxd 3/8/04 11:09 AM Page 58
58
Chapter 3
■■
Change the mix of acquisition channels to favor those that bring in the most loyal customers.
■■
Forecast customer population in future months.
■■
Alter the product to address defects that are causing customers to defect.
Each of these goals has implications for the data mining process. Contacting existing customers through an outbound telemarketing or direct mail campaign implies that in addition to identifying customers at risk, there is an understanding of why they are at risk so an attractive offer can be constructed, and when they are at risk so the call is not made too early or too late. Forecasting implies that in addition to identifying which current customers are likely to leave, it is possible to determine how many new customers will be added and how long they are likely to stay. This latter problem of forecasting new customer starts is typically embedded in business goals and budgets, and is not usually a predictive modeling problem.
How Will the Results Be Delivered?
A data mining project may result in several very different types of deliverables. When the primary goal of the project is to gain insight, the deliverable is often a report or presentation filled with charts and graphs. When the project is a one-time proof-of-concept or pilot project, the deliverable may consist of lists of customers who will receive different treatments in a marketing experiment. When the data mining project is part of an ongoing analytic customer relationship management effort, the deliverable is likely to be a computer program or set of programs that can be run on a regular basis to score a defined subset of the customer population along with additional software to manage models and scores over time. The form of the deliverable can affect the data mining results. Producing a list of customers for a marketing test is not sufficient if the goal is to dazzle marketing managers.
The Role of Business Users and Information Technology
As described in Chapter 2, the only way to get good answers to the questions posed above is to involve the owners of the business problem in figuring out how data mining results will be used and IT staff and database administrators in figuring out how the results should be delivered. It is often useful to get input from a broad spectrum within the organization and, where appropriate, outside it as well. We suggest getting representatives from the various constituencies within the enterprise together in one place, rather than interviewing them separately. That way, people with different areas of knowledge and expertise have a chance to react to each other’s ideas. The goal of all this consultation is a clear statement of the business problem to be addressed. The final
470643 c03.qxd 3/8/04 11:09 AM Page 59
Data Mining Methodology and Best Practices
59
MISUNDERSTANDING THE BUSINESS PROBLEM: A CAUTIONARY TALE
Data Miners, the consultancy started by the authors, was once called upon to analyze supermarket loyalty card data on behalf of a large consumer packaged goods manufacturer. To put this story in context, it helps to know a little bit about the supermarket business. In general, a supermarket does not care whether a customer buys Coke or Pepsi (unless one brand happens to be on a special deal that temporarily gives it a better margin), so long as the customer purchases soft drinks. Product manufacturers, who care very much which brands are sold, vie for the opportunity to manage whole categories in the stores. As category managers, they have some control over how their own products and those of their competitors are merchandised. Our client wanted to demonstrate its ability to utilize loyalty card data to improve category management. The category picked for the demonstration was yogurt because by supermarket standards, yogurt is a fairly high-margin product.
As we understood it, the business goal was to identify yogurt lovers. To create a target variable, we divided loyalty card customers into groups of high, medium, and low yogurt affinity based on their total yogurt purchases over the course of a year and into groups of high, medium, and low users based on the proportion of their shopping dollars spent on yogurt. People who were in the high category by both measures were labeled as yogurt lovers.
The transaction data had to undergo many transformations to be turned into a customer signature. Input variables included the proportion of trips and of dollars spent at various times of day and in various categories, shopping frequency, average order size, and other behavioral variables.
Using this data, we built a model that gave all customers a yogurt lover score.
Armed with such a score, it would be possible to print coupons for yogurt when likely yogurt lovers checked out, even if they did not purchase any yogurt on that trip. The model might even identify good prospects who had not yet gotten in touch with their inner yogurt lover, but might if prompted with a coupon.
The model got good lift, and we were pleased with it. The client, however, was disappointed. “But, who is the yogurt lover?” asked the client. “Someone who gets a high score from the yogurt lover model” was not considered a good answer. The client was looking for something like “The yogurt lover is a woman between the ages of X and Y living in a zip code where the median home price is between M and N.” A description like that could be used for deciding where to buy advertising and how to shape the creative content of ads. Ours, based on shopping behavior rather than demographics, could not.
statement of the business problem should be as specific as possible. “Identify the 10,000 gold-level customers most likely to defect within the next 60 days”
is better than “provide a churn score for all customers.”
The role of the data miner in these discussions is to ensure that the final statement of the business problem is one that can be translated into a data mining problem. Otherwise, the best data mining efforts in the world may be addressing the wrong business problem.
470643 c03.qxd 3/8/04 11:09 AM Page 60
60
Chapter 3
Data mining is often presented as a technical problem of finding a model that explains the relationship of a target variable to a group of input variables.
That technical task is indeed central to most data mining efforts, but it should not be attempted until the target variable has been properly defined and the appropriate input variables identified. That, in turn, depends on a good understanding of the business problem to be addressed. As the story in the sidebar illustrates, failure to properly translate the business problem into a data mining problem leads to one of the dangers we are trying to avoid—
learning things that are true, but not useful.
For a complete treatment of turning business problems into data mining problems, we recommend the book Business Modeling and Data Mining by our colleague Dorian Pyle. This book gives detailed advice on how to find the business problems where data mining provides the most benefit and how to formulate those problems for mining. Here, we simply remind the reader to consider two important questions before beginning the actual data mining process: How will the results be used? And, in what form will the results be delivered? The answer to the first question goes a long way towards answering the second.
Step Two: Select Appropriate Data
Data mining requires data. In the best of all possible worlds, the required data would already be resident in a corporate data warehouse, cleansed, available, historically accurate, and frequently updated. In fact, it is more often scattered in a variety of operational systems in incompatible formats on computers running different operating systems, accessed through incompatible desktop tools.
The data sources that are useful and available vary, of course, from problem to problem and industry to industry. Some examples of useful data:
■■
Warranty claims data (including both fixed-format and free-text fields)
■■
Point-of-sale data (including ring codes, coupons proffered, discounts applied)
■■
Credit card charge records
■■
Medical insurance claims data
■■
Web log data
■■
E-commerce server application logs
■■
Direct mail response records
■■
Call-center records, including memos written by the call-center reps
■■
Printing press run records
470643 c03.qxd 3/8/04 11:09 AM Page 61
Data Mining Methodology and Best Practices