Logistic Model Trees with AUC Split Criterion for the KDD Cup 2009 Small Challenge

In this work, we describe our approach to the “Small Challenge” of the KDD cup 2009, the prediction of three aspects of customer behavior for a telecommunications service provider. Our most successful method was a Logistic Model Tree with AUC as split criterion using predictions from boosted decision stumps as features. This was the best submission for the “Small Challenge” that did not use additional data from other feature sets. A second approach using an AUC-optimized weighted linear combination of several rankings scored slightly worse with an average AUC of 0.8074. From the given 230 features, we extracted additional binary features and imputed missing values using SVMs and Decision Trees.