An adaptive Bagging algorithm based on lightweight transformer for multi-class imbalance recognition

The class imbalance is a significant issue in machine learning, particularly in the context of multi-class imbalance. The current multi-class imbalanced classifiers often encounter the risk of over-fitting and poor scalability. We propose an imbalanced Bagging framework which could be combined with lightweight to maintain high classification accuracy and low computational complexity. In addition, we introduce an adaptive undersampling method for multi-learner system, which could adaptively provide sampling schemes based on category proportions and significantly reduce processing time and over-fitting. Finally, we propose a novel adaptive cost-sensitive loss function for highly imbalanced scenarios, which could improve the robustness of the model in assigning weights to class samples and enhance the recognition accuracy of minority class samples. It is worth noting that our proposed method has good scalability and could achieve state-of-the-art on some data sets by combining with some algorithms. The experiments are conducted on multiple popular imbalanced data sets, including cifar10-LT, cifar100-LT and Place365-LT, and the models are evaluated in terms of multiple imbalanced metrics. The comparative experimental results show that the accuracy of the proposed method is superior to the current state-of-the-art method on imbalanced data sets. The ablation experiments demonstrate the effectiveness and progressiveness of our proposed approach. The proposed method has universality and could be easily transferred to tasks in other multimedia systems.