Boosting first-order clauses for large, skewed data sets

Abstract

Creating an effective ensemble of clauses for large, skewed data sets requires finding a diverse, high-scoring set of clauses and then combining them in such a way as to maximize predictive performance. We have adapted the RankBoost algorithm in order to maximize area under the recall-precision curve, a much better metric when working with highly skewed data sets than ROC curves. We have also explored a range of possibilities for the weak hypotheses used by our modified RankBoost algorithm beyond using individual clauses. We provide results on four large, skewed data sets showing that our modified RankBoost algorithm outperforms the original on area under the recall-precision curves. © 2010 Springer-Verlag Berlin Heidelberg.