In the advertising world, we need to predict who will do the action for which we get rewarded (“convert”) after we serve them an ad. Conversion is fairly rare, so we need a robust approach to predicting rare events. We call a list of users, ordered by propensity to convert, an “audience”. In this talk I will describe a model ensemble method that can be used to build an audience: stacking. Stacking not only strengthens the final model by reducing model bias, it also allows parallelization of model development amongst developers. In addition to describing the model itself and how it relates to other ensemble models, I will address how to optimize a stacked model ensemble and pitfalls to avoid. I will
describe how to implement k-fold cross validation on a stacked ensemble of models to avoid bias. At the end of this talk, the audience will know what stacking is and why it is useful in the context of audience modeling. They will have been introduced to best practices for stacking implementation and tuning, both from the perspective of the algorithm development as well as software development in python.
Alice Broadhead has been a Data Scientist at Valassis Digital for the past four years. Her work has provided fundamental research and development for new geolocation, foot traffic monitoring and audience selection capabilities that are core to the Valassis Digital managed services. Alice has a bachelor’s degree in mathematics and biology and a PhD in plant biology. She sees herself as a recovering tree physiologist who can now see the Random Forests for the trees.