A Composite Cost-Sensitive Neural Network for Imbalanced Classification

Abstract

Imbalanced data with skewed class distributions and different misclassification costs is common in many real-world applications. Traditional classification approach does not work well for imbalanced data, because they assume equal costs for each class. To deal with this problem, cost-sensitive approaches assign different misclassification costs for different classes without disrupting the true original distributions of samples. However, due to lack of prior knowledge, the misclassification costs are usually unknown and hard to choose in practice. Whats more, even instances in the same class may have different misclassification costs. As an extension of class-dependent costs, this paper presents a composite cost-sensitive deep neural network (CCS-DNN) for imbalanced classification. A specifically-designed cost-sensitive matrix, which is composed of example-dependent costs and class-dependent costs, is embedded into the loss function to improve the classification performance. And the parameters of both the cost-sensitive matrix and the network are jointly optimized during training. The results of comparative experiments on some benchmark datasets indicate that the CCS-DNN performs better than other baseline methods.