Klimasauskas Group^TM
Leading the Edge in Emerging Technologies

Marketing applications - targeting the specific group of customers most likely to respond to your offering - is a well-developed industry. This is one of the primary targets of the "Data Mining" industry. Traditional uses of statistics are in the direct mail arena. More recently, with the explosion of the internet, focus has turned to click-stream analysis, and demographic profiling. Klimasauskas Group has worked with a range of applications including direct mail and dynamic on-line profiling for web-sites.

Direct Mail. Staff of the Klimasauskas Group have worked with several direct marketers in developing predictive models for direct mail. In support of these applications, we have developed several technologies that represent breakthroughs in developing high performance models in less time. Three areas in particular are: 1) Automatic data transformation; 2) Genetic variable selection; 3) Data conflict resolution for improving overall model performance. Automatic data transformation provides mechanisms for analyzing the distribution of input data and developing a series of potential transformations that make the distribution more nearly uniform. Several candidate transformations are generated for each input variable. Genetic algorithms are used to select synergistic subsets of transformed variables. These synergistic subsets can be used to train multiple models which "vote" together. Our research has shown that these synergistic subsets of data achieve their performance by looking at the data from different perspectives. Combining them together through different models in a "voting" block can significantly reduce false positives. Data conflict resolution uses the observation that several modeling technologies - including logistic regression - will find sub-optimal solutions (local minima) when developed using conflicting data (multiple input vectors that are nearly identical with conflicting outcomes). Data conflict resolution applies a series of algorithms to the data to divide the space into independent sub-cubes. A representative sample is computed for each non-null sub-cube. The reduced data set is used for model development. Whether developing logistic regression models, or neural network models, we have found that this technique produces more stable solutions, and finds them with less computational resources. It is our experience that a well-selected, effectively transformed set of variables, used to train a logistic regression model on a conflict-reduced data set, often will achieve similar or better performance than many Neural Network approaches. If you are involved in developing statistical or Neural Network models for direct mail, and want to improve your systems performance, contacts us below.

Dynamic On-line Profiling. This conceptual proto-type was developed for a company doing audience profiling for television. The basic idea is that you want to be able to monitor viewing behavior, and from that determine a demographic profile of the current viewer so that you can decide which advertisements to show during the intermissions. The challenge is that several individuals may watch the same television set. What you would like to do is to identify the dominant viewer by the most recent program or two that they have watched, and refine their profile over time. This was accomplished by using concepts from Neural Networks to form demographic clusters based on programs watched. Through statistical analysis, it was possible to identify when a new cluster should be created. For any given program, the cluster closest to the standard demographic profile for that program is selected, and a probability distribution associated with that profile updated based on the current program. This application illustrates the capabilities of Klimasauskas Group staff to apply a range of technologies to novel problems to deliver practical solutions.

Key Benefits

• Improved direct mail response rate.
• Simple models similar to those developed with statistical methods.
• Effective demographic identification.

Capabilities

These applications illustrate some of the capabilities of Klimasauskas Group. Even though Neural Network technology has become more readily available in the various statistical software packages available, and in many of the Data Mining suites, there is still a reluctance to utilize black-box technology that can't be understood or analyzed. We offer an alternative - the ability to use techniques inspired by Neural Network technologies to improve existing logistic regression and multiple-logistic regression models. If this sounds interesting to you, contact us below.