Analysis of the effect of class imbalance on transfer learning

Class imbalance of datasets is a common problem in the field of machine learning and transfer learning is no exception. However, very limited research is available about the effect of class imbalance on transfer learning, this paper focuses on the analysis of the effects of several imbalanced classification algorithms on transfer learning to address the issue: oversampling, undersampling, weighted random sampling, weighted cross entropy loss, Focal Loss and L2RW algorithm based on meta learning. Among them, the first three methods eliminate the imbalance of the dataset by random sampling, weighted cross entropy loss and Focal Loss keep the dataset unchanged and adjust the loss function of standard classification algorithms, and L2RW algorithm adopts meta learning mechanism to adjust the weight of training set sample dynamically to achieve better performance in generalization. Extensive empirical evidence shows that oversampling and weighted random sampling are more suitable for transfer learning among various imbalanced classification algorithms.