A Trajectory Unified Framework for Deep Neural Network Training in Imbalanced Datasets

This paper proposes a novel optimization method designed to address the challenges of training deep neural networks on imbalanced datasets. The main contributions include the proposal of a novel trajectory-unified method based on the influence balance factor, which adaptively reweights samples to alleviate the overfitting of decision boundaries to the majority class. In this paper a rigorous theoretical framework is presented that validates the stability and convergence properties of the proposed method. Extensive experiments demonstrate that this method outperforms cutting-edge techniques in handling class imbalance issues. In summary, this approach effectively tackles the challenges posed by imbalanced data distribution, providing a systematically enhanced solution to the class imbalance problem in advanced deep learning models.