BGDSO: A New Synthetic Oversampling Method Based on Bilateral Gradient Distribution for Imbalanced Data

For the optimization problem of high-dimensional imbalanced data, we propose a dual optimization theoretical framework that integrates global gradient and fine-grained gradient distribution balance to further improve the limitations of current methods that only rely on the global gradient distribution of minority classes. We pointed out that the existing methods do not make full use of the gradient distribution information of the majority class samples, and optimize the class contribution balance only from the global perspective, ignoring the local balance problem within the fine-grained gradient interval. For this reason, the concept of bilateral gradient distribution (BGD) is innovatively proposed and a synthetic oversampling method of bilateral gradient distribution (BGDSO) is designed based on it. The method ensures that the synthetic samples approximate the gradient distribution of the original minority class while requiring them to be consistent with the distribution of the majority class in the safe gradient interval, so as to realize the dual balance of global and local fine-grained gradients. The superiority of the algorithm is verified by experiments on real data sets, and the generalization performance test and statistical test of the algorithm are also implemented. The results demonstrating its theoretical value and practical significance in improving the classification performance of high-dimensional imbalanced data.