In deep learning contexts, model performance often degrades significantly when trained on heavily imbalanced datasets, particularly when the evaluation metrics necessitate robust generalization across infrequently represented classes. In addressing the challenges posed by imbalanced data distributions in machine learning, this study introduces a novel approach that employs the method of density ratio estimation for dynamic class weight adjustment during model training, an innovative method we refer to as Re-weighting with Density Ratio (RDR). Our approach enables real-time adjustment of the importance of each class during the training process, mitigating overfitting on majority classes and enhancing adaptability across diverse datasets. Extensive experiments conducted on various large scale benchmark datasets validate the effectiveness of our approach. Results demonstrate substantial improvements in generalization capabilities, particularly under severely imbalanced condition.