When Predictive Models Join Forces: on the How and Why of Ensemble Learning for Customer Intelligence/ Database Marketing (By Koen W. De Bock, PhD)
In customer intelligence, predictive modeling is a key instrument. Applications such as customer churn prediction, response modeling, cross-sell (Next-Product-To-Buy) analysis or customer lifetime value analysis all depend upon inferences about expected future customer behavior or characteristics in order to make marketing campaigns more targeted and effective. While the success of these activities depends on decisions made in several phases of the modeling process, it is widely acknowledged that the choice of the modeling technique is a prominent one. In this presentation, light is shed upon the advantages of letting predictive models in CI join forces, whereby several models are combined into new and more powerful models. These so-called ensembles have consistently emerged as winning entries in data mining contests, such as the Teradata/Duke CRM competition, KDD Cup or the Netflix Prize since many years. However, despite their strength and intuitive nature, their applications in real-life business are still scarce. This talk will untangle the topic of ensemble learning, include an overview of the most important techniques, how they can be tailored to get the most out of your CI applications. The advantages of the techniques are demonstrated throughout several academic experiments on real-life datasets.
In other words, the model will have severe difficulties to identify the heartbreaking customers. If you don’t do anything about this problem, the accuracy of your model will seem extremely good, but actually your model will not be able to make a good distinction between churners and non-churning customers. 
