Class Imbalanced Semi-Supervised Learning With Meta-Learning
Yunyi Shi, Ming Zhang, Xin Zheng, Hongyu Chen · 2023
In recent years, in order to solve the lack of labeled data faced by deep learning, semi-supervised learning (SSL) algorithms based on pseudo-label methods and consistency regularization methods have achieved remarkable results. These algorithms can achieve performance comparable to supervised learning with only a small amount of data labels, but they perform poorly on class-imbalanced datasets in practical applications. On the one hand, the class imbalance problem will cause the model classification to shift towards the majority class, and on the other hand, it will cause pseudo-labels noise generated by the SSL algorithm. These two challenges together harm the performance of these algorithms on class-imbalanced datasets. We propose a meta sample reweighting algorithm to simultaneously address the above two challenges. We introduce two meta networks to adaptively weight labeled and unlabeled data respectively, use meta-learning framework to learn the parameters of the meta networks, and propose a meta learning algorithm that alternately updates the parameters of the meta-network and the classifier network. Experiments prove that this method can effectively alleviate the performance degradation of state-of-the-art SSL algorithms affected by the class imbalance problem.