Adaptive weight optimization for classification of imbalanced data

One popular approach for imbalance learning is weighting samples in rare classes with high cost and then applying cost-sensitive learning methods to deal with imbalance in classes. Weight of a class is usually determined by proportion of samples in each class in training set. This paper analyzes that sample proportions of training set and testing set may vary in some range and it would compromise performance of learned classifier. This problem becomes serious when class distribution is extremely high imbalanced. Based on the analysis, an adaptive weighting approach aiming at finding a group of proper weights for classes is proposed. We employ evolutionary algorithm to optimize weight configuration to ensure overall performance of classifier in both training set and possible testing sets. Experimental results on a wide variety of datasets demonstrate that our approach could achieve better performances. ? 2013 Springer-Verlag Berlin Heidelberg.EI