|
Abstract:
|
Segmentation classifiers, or "typing tools," are models for predicting customer memberships in existing segmentation schemes. The prediction task can be accomplished with any of a wide variety of methods. Recent advances in the development of "deep neural networks" (DNN's) provide a new alternative. Their features include being able to model high level abstractions of data configurations in order to maximize predictive accuracy, having some methods for mitigating overfitting, and being widely deployable in many forms. DNN challenges include making the large number of decisions needed to specify their topologies and learning algorithm(s), difficulty in interpreting what is usually a very large number of parameter estimates, and a general lack of well understood prediction error distributions. But they can be useful. We explain how DNN's can be used for segment classification. We describe their application to real customer data exemplary of the type encountered in conventional targeting applications. Lastly, we summarize open source resources for fitting, evaluating, and applying DNN's.
|
