Decoupling Deep Domain Adaptation Method for Class-imbalanced Learning with Domain Discrepancy

In a wide range of classification tasks, training data will produce class-imbalance due to collection difficulties in some classes, which leads to prediction biases on minority classes. For the class-imbalanced problem, existing researches are usually based on the assumption that the training dataset and the test dataset are from similar distributions. In reality, both of the datasets often come from domains with different distributions, which challenges generalization performances of models. In this paper, a decoupling deep domain adaptation method is proposed to overcome these problems. Based on the adversarial domain adaptation model, the method uses a two-stage training strategy which decouples representation learning and classifier adjustment. The results of experiments under scenarios of bearing fault diagnosis and digit images classification with class-imbalance and domain discrepancy show that the effect of the combination of domain adaptation method and specific decoupling strategy is better than that of one-stage training only using resampling or cost-sensitive methods in the domain adaptation model.

1 Introduction

Class-imbalance is an important problem in the research of data-driven artificial intelligence. Limited by various conditions, data collection in many fields often cannot obtain enough equal samples in all classes. For example, in the research of industrial fault diagnosis, because the rotating machinery works normally most of the time, the number of fault samples that can be collected is far less than that of normal samples [1]. In addition, in the field of image classification, there are many inevitable image class-imbalanced problems in the real world, such as medical image analysis, anomaly detection, disaster prediction and so on [2]. Class-imbalance will lead to model prediction biases towards majority classes [3].