DuNet: A Robust End-to-End Deep Neural Network Framework for Imbalanced Classification

Abstract

Imbalanced classification is a prevalent task in machine learning, characterized by unequal sample distribution among classes. The majority of classifiers aim to optimize prediction performance across all samples, leading to a bias toward the majority class. In this paper, we propose an end-to-end deep neural network framework called DuNet, specifically designed for the imbalanced classification task. In DuNet, we design a module known as the Dual Auto-Encoder (DAE), which adeptly generates discriminative features to differentiate between majority and minority classes. DuNet demonstrates significant superiority over other baseline models and achieves state-of-the-art results on real-world imbalanced classification datasets and the Breast Cancer Wisconsin (Diagnostic) medical dataset (WDBC).