Multi-class Boosting for Imbalanced Data

We consider the problem of multi-class classification with imbalanced data-sets. To this end, we introduce a cost-sensitive multi-class Boosting algorithm (BAdaCost) based on a generalization of the Boosting margin, termed multi-class cost-sensitive margin. To address the class imbalance we introduce a cost matrix that weighs more hevily the costs of confused classes and a procedure to estimate these costs from the confusion matrix of a standard 0 vertical bar 1-loss classifier. Finally, we evaluate the performance of the approach with synthetic and real datasets and compare our results with the AdaC2.M1 algorithm.