An Empirical Boosting Scheme for ROC-Based Genetic Programming Classifiers

The so-called “boosting” principle was introduced by Schapire and Freund in the 1990s in relation to weak learners in the Probably Approximately Correct computational learning framework. Another practice that has developed in recent years consists in assessing the quality of evolutionary or genetic classifiers with Receiver Operating Characteristics (ROC) curves. Following the RankBoost algorithm by Freund et al., this article is a cross-bridge between these two techniques, and deals about boosting ROC-based genetic programming classifiers. Updating the weights after a boosting round turns to be the algorithm keystone since the ROC curve does not allow to know directly which training cases are learned or misclassified. We propose a geometrical interpretation of the ROC curve to attribute an error measure to every training case. We validate our ROCboost algorithm on several benchmarks from the UCI-Irvine repository, and we compare boosted Genetic Programming performance with published results on ROC-based Evolution Strategies and Support Vector Machines.