In addition to their success in many Kaggle competitions, at Civis Analytics we have found that stacked models produce reliable push-button estimators for many of the problems we encounter. To enable the easy creation of stacked classifiers and regressors in python, we've written a pair of scikit-learn-compatible meta-estimators that handle both the the I/O between the layers of models and the cross-validation needed for fitting the full model stack.

This talk will describe the methodology of model stacking, along with best practices. We'll show how our stacked model meta-estimators work under the hood. We will also demonstrate how to construct stacked models yourself using the open-source code described above. The source code for our model stacking meta-estimators can be found on github. This code can be installed with pip by calling pip install civisml-extensions.