GEV-NN: A deep neural network architecture for class imbalance problem in binary classification

Class imbalance is a common issue in many applications such as medical diagnosis, fraud detection, web advertising, etc. Although standard deep learning method has achieved remarkably high-performance on datasets with balanced classes, its ability to classify imbalanced dataset is still limited. This paper proposes a novel end-to-end deep neural network architecture and adopts Gumbel distribution as an activation function in neural networks for class imbalance problem in the application of binary classification. Our proposed architecture, named GEV-NN, consists of three components: the first component serves to score input variables to determine a set of suitable input, the second component is an auto-encoder that learns efficient explanatory features for the minority class, and in the last component, the combination of the scored input and extracted features are then used to make the final prediction. We jointly optimize these components in an end-to-end training. Extensive experiments using real-world imbalanced datasets showed that GEV-NN significantly outperforms the state-of-the-art baselines by around 2% at most. In addition, the GEV-NN gives a beneficial advantage to interpret variable importance. We find key risk factors for hypertension, which are consistent with other scientific researches, using the first component of GEV-NN. (C) 2020 Elsevier B.V. All rights reserved.