David Raab Article Archive

Recognition Systems Ideas
David M. Raab
DM News
March, 1998

Two years ago, anybody who wanted to run database marketing campaigns on “merchant” relational databases like Oracle and Sybase had only a few pioneering products to choose from. Today, there are more than two dozen “open systems” campaign managers stands–of which half were been introduced in the past year. With this sudden explosion in the market, systems that would have been extraordinary in the past now seem barely worth notice. Only products with capabilities significantly beyond market leaders like Exchange Applications and Prime Response can hope to attract detailed attention.

Ideas (Recognition Systems Inc., 312-382-8989, www.recsys.com) attempts to distinguish itself with unprecedented integration of predictive models and campaign development. It does this by building models with an automated model-builder, displaying model results on a graphical gains chart, and letting users define campaign segments by highlighting bands of records on the chart. Ideas will automatically generate the SQL code to select these segments during the campaign execution process.

This is considerably more streamlined than the way most systems incorporate predictive models, which usually involves building the models separately, importing either the scoring algorithms or scored records themselves, and creating segment definitions that include score ranges. By removing nearly all the technical administration from the process, Ideas hopes to allow marketers to incorporate models into nearly all of their projects.

It turns out that removing the technical roadblocks involves changes to a great many tasks. As experienced model builders know, most of the work revolves around data preparation–extracting and transforming the raw input–rather than the modeling itself. Ideas provides an elaborate data model to make marketing data easily accessible. This model treats all customer-related data as transactions, including items such as demographics that are often just replaced without keeping track of prior values. The system can automatically transform the raw transactions into summary and current views of the data, so that all three levels (detail, summary and current) are available without special preparation by the end-user.

The data model itself is unusually complete. In addition to the usual purchase history and communications history, it stores information about competitive promotions available in a given market at a give time, so these can also be factored into models and response analysis. The customer level of the model also includes marketing budgets and target numbers of contacts for individual customers, to help marketers allocate resources on a customer-by-customer basis. The system does not determine the optimal budget or contact levels automatically, although its modeling abilities can help end-users with this analysis.

The model-building components are competitive with other “end-user” modeling systems available today. They can automatically split the data into training and validation samples, prepare and transform the data, identify appropriate variables, build and compare multiple models using several regression and neural network techniques, display results on the gains chart, and provide evaluation statistics on error rates and variable importance. Users define the value to predict and the specific goal of the model–for example, whether to maximize discrimination on the top 20% of the file or for the entire universe. Advanced users have the option to specify the variables and modeling techniques to be used. Ideas can also use results from a “committee” of several different models in the final record scoring. The vendor plans to add decision tree (CHAID or CART) models some time this year.

Model building is done on the user’s workstation using a sample of data extracted from the central database. Campaign development and reporting are also done on the local workstation. The system automatically goes back to the central computer when the user requests data that has not already been stored locally.

Campaigns are built by first creating a selection tree to define the file segments and then defining the promotions sent to each segment. Segments in the selection tree can be based on model scores, a table of values joined by simple and/or logic, or user-written SQL code for more complicated conditions. The contents of one segment can be passed to another segment to be split further, and the contents of several segments can be combined in a new segment. The system does not automatically eliminate duplicates across segments, although the vendor plans to add this as an option. Ideas will generate segment counts on demand from the user.

Once the segments are defined, users can attach “objects” that control

list generation, reports, models, extraction to Excel spreadsheets, costs, revenues and execution schedules. Like the data model and integrated modeling, objects let end-users perform tasks that would otherwise require intervention by technical support staff.

The list generation object attaches a 5 x 5 matrix of offers and treatments to each segment. For each cell in the matrix, users can allocate the quantity or percentage of names, assign keycodes, store revenue and cost work sheets (which are themselves objects), and specify outputs. The system will randomly allocate names among cells in matrix. If a model has been used to create the segment, the system will use the model results to forecast response quantity–a nice feature, although the forecast will be the same for all offer/treatment combinations unless these have been modeled separately. The promotion object can store information about each offer and treatment, but the system does not maintain a list of standard offers and treatments that can be used in different promotions. Templates help users set up promotions with test/control structures and response analysis reports. Reporting is done with objects that extract data to Excel spreadsheets, using templates supplied with the system or added by the user.

Ideas runs on Windows NT workstations. It uses ODBC to access data on NT, Unix or mainframe servers. Pricing is based on the size of the purchasing company and starts at about $350,000. The system is built to work with existing marketing databases and does not provide tools for database development or maintenance. It was introduced in 1997 and currently has five installations.

* * *