DECISION TECHNOLOGIES IN DATABASE MARKETING: PART IV
by Gene M. Ferruzza, Senior VP, Decision Technologies

Matching at the customer level is critical for assembling a data profile for each individual customer in the data mart. In a data profile, all customer activity and information, across all of the data mart's data sources, should be accessible through one unique customer identifier. The variants of addresses become even more diverse when records are matched for householding applications, because data collected on the household come from different individuals in the same household.

Customer matching and householding contribute to a company's understanding of the customers in the data mart, and address standardization improves this process. It also enables the user to append additional data at the customer (individual) or household level, because addresses must be standardized in order to be matched with those of the appended data.

Some may question whether something as seemingly mundane as name and address matching qualifies as a data-mining process. However, as a process that involves pattern recognition, it is similar to other functions more widely recognized as data mining. For example, the use of patterns of data from names and addresses to identify unique customers and customers who reside in the same household is similar to the use of patterns of data to identify customers who will respond to an offer (based on patterns of data from customers who have responded to the same offers in the past).

Both response-modeling processes and matching and de-duplication processes use pattern-recognition algorithms. When we standardize an address or merge two records, we have, in effect, created new, useful information. It is a common practice to acquire data from third parties that cannot be obtained through usual customer relationships and append these third-party data to prospects' or customers' records. By appending demographic or psychographic data to customer records, companies widen the scope of data mining opportunities. The transformation of a list of names and addresses into an information-rich data source is especially useful for many marketing programs. Combining this information with customer behavior data can provide a powerful foundation for the use of data mining.

Basic and Complex Data Mining

Basic data-mining approaches organize and filter the data so that knowledge is uncovered and becomes apparent to the user. The route to information often is known, and the data operations to get there are clear. Basic approaches are analytical operations in the form of queries, summarization operations, simple statistics, or fixed formulas. These techniques are relatively easy to develop and understand, and they usually are designed by users having expertise with the customer data and the business application (i.e., domain experts). A common term for such basic data mining is "data analysis," which in reality is no more descriptive a term than "data mining."

It is important to have an expert in the business application who can ask the right questions of the data (i.e., construct the proper queries). With user-friendly and sophisticated software, business users may conduct their own queries of the data; otherwise, they should work closely with a database specialist or data miner. In many cases, a formula already exists, based on the experiences of one or more domain experts.

Basic data-mining approaches are the most popular and useful approaches for marketing applications as long as domain expertise exists. However, even experts usually are not capable of formulating queries that will accurately forecast customer behavior such as attrition or response. Thus, it is difficult, if not impossible, to accurately predict customer behavior with basic approaches.

In cases where domain expertise does not exist or where customer behavior is complex, data miners refer to the data to provide information to drive marketing programs using what are referred to as "complex approaches." In contrast to basic data-mining approaches, complex approaches are data-driven; in other words, patterns of data in the database are used to construct algorithms based on target variables that also reside in the database. For instance, a customer response model uses historical patterns of customer behavior (as represented in the data) to predict a customer's future response. The model is developed through the use of customers' past responses or non-responses as a target (or dependent) variable. The resulting model maps current customer data patterns to a probability of response.

Part V of this series will appear in the next edition of D S * .

Contact Gene Ferruzza at gmf@cmsnet.com