Abstaining in rule set bagging for imbalanced data
Krystyna Napierała, Jerzy Stefanowski · Logic Journal of IGPL · 2015
Class imbalanced data constitute difficulties for most classifiers. Standard ensembles fail to sufficiently recognize the minority class. The role of classification strategies in rule set based component classifiers inside bagging is studied. We argue that introducing abstaining in bagging, i.e. allowing component classifiers to refrain from predicting class labels in ambiguous situations, improves classification of imbalanced data. Comparative experiments with 5 different strategies and 2 rule induction algorithms confirm this hypothesis.