An Ensemble Pruning Approach Based on Reinforcement Learning in Presence of Multi-class Imbalanced Data

Abstract

In recent years, learning from imbalanced data sets has become a challenging issue in machine learning and data mining communities. This problem occurs when some classes of data have smaller number of instances than other classes. Multi-class imbalanced data sets have been pervasively observed in many real world applications. Many typical machine learning algorithms pose many difficulties dealing with these kinds of data sets. In this paper, we proposed an ensemble pruning approach which is based on Reinforcement Learning framework. In effect, we were inspired by Markov Decision Process and considered the ensemble pruning problem as a one player game, and select the best classifiers among our initial state space. These selected classifiers which can produce a good ensemble model, are employed to learn from multi-class imbalanced data sets. Our experimental results on some UCI and KEEL benchmark data sets show promising improvements in terms of minority class recall, G-mean, and MAUC.