Adapting MultiBoost ensemble for class imbalanced learning

Learning with class imbalanced data sets is a challenging undertaking by the common learning algorithms. These algorithms favor majority class due to imbalanced class representation, noise and their inability to expand the boundaries of minority class in concept space. To improve the performance of minority class identification, ensembles combined with data resampling techniques have gained much popularity. However, these ensembles attain higher minority class performance at the cost of majority class performance. In this paper, we adapt the MultiBoost ensemble to deal with the minority class identification problem. Our technique inherits the power of its constituent and therefore improves the prediction performance of the minority class by expanding the concept space and overall classification performance by reducing bias and variance in the error. We compared our technique with seven existing simple and ensemble techniques using thirteen data sets. The experimental results show that proposed technique gains significant performance improvement on all tested metrics. Furthermore, it also has inherited advantage over other ensembles due to its capability of parallel computation.