Unbalanced Data Integration Learning Based on Kernel Density Estimation

To effectively address the challenges of undersampling techniques when handling imbalanced data, a new undersampling ensemble learning algorithm based on Kernel Density Estimation (KDEE) is proposed and designed. The algorithm employs a novel binning mechanism to coordinate the imbalance rate of the data through bin sampling, generating high-quality training subsets for iterative classifier training. During the iteration process, weights are assigned to base classifiers based on their G-mean score on the original data, allowing classifiers with better performance to have a greater impact on the overall decision, thereby improving the generalization ability of the ensemble model. The new algorithm increases the visibility of the minority class and avoids the impact of class overlap. Comparative experiments with small and large datasets (4 groups) demonstrate that the algorithm has advantages in handling imbalanced data classification problems.