A comparative evaluation of feature set evolution strategies for multirelational boosting

Abstract

Boosting has established itself as a successful technique for decreasing the generalization error of classification learners by basing predictions on ensembles of hypotheses. While previous research has shown that this technique can be made to work efficiently even in the context of multirelational learning by using simple learners and active feature selection, such approaches have relied on simple and static methods of determining feature selection ordering a priori and adding features only in a forward manner. In this paper, we investigate whether the distributional information present in boosting can usefully be exploited in the course of learning to reweight features and in fact even to dynamically adapt the feature set by adding the currently most relevant features and removing those that are no longer needed. Preliminary results show that these more informed feature set evolution strategies surprisingly have mixed effects on the number of features ultimately used in the ensemble, and on the resulting classification accuracy.