Enhancing the Performance of Multiclass CSMLP on Imbalanced Data Using AdaBoost.NC Algorithm

Abstract

Cost-sensitive methods have a strong theoretical appeal, since they explicitly violate the equality principle of traditional error-based learning algorithms. Possibility of incorporating prior information via individual loss functions (costs) may lead to unbiased classification models, which improve the detection of under-represented classes. Proposed Cost function that is based on the linear combination of individual objectives associated to each class. This simplified formulation enables the assignment of unequal misclassification costs from a single cost parameter. Levenberg?Marquadt (LM) learning method to that gradient vectors and Hessian matrices can be calculated separately for each class. Besides allowing learning with asymmetric costs, the proposed approach retains the efficiency of the Gauss?Newton optimization method. The extension of the Levenberg?Marquadt learning rule which is applied to Multi Layer Perception (MLP) increases the accuracy of the classifier. In addition, it is theoretically demonstrated that a balanced decision boundaries in the feature space is achieved by incorporating of prior information via the cost parameter.