Evolving Neural Networks with Maximum AUC for Imbalanced Data Classification

Real-world classification problems usually involve imbalanced data sets. In such cases, a classifier with high classification accuracy does not necessarily imply a good classification performance for all classes. The Area Under the ROC Curve (AUC) has been recognized as a more appropriate performance indicator in such cases. Quite a few methods have been developed to design classifiers with the maximum AUC. In the context of Neural Networks (NNs), however, it is usually an approximation of AUC rather than the exact AUC itself that is maximized, because AUC is non-differentiable and cannot be directly maximized by gradient-based methods. In this paper, we propose to use evolutionary algorithms to train NNs with the maximum AUC. The proposed method employs AUC as the objective function. An evolutionary algorithm, namely the Self-adaptive Differential Evolution with Neighborhood Search (SaNSDE) algorithm, is used to optimize the weights of NNs with respect to AUC. Empirical studies on 19 binary and multi-class imbalanced data sets show that the proposed evolutionary AUC maximization (EAM) method can train NN with larger AUC than existing methods.