Safety adaptive oversampling method combined with radial-based undersampling
Chenxiao Zhou · 2024
Sample imbalance is an important problem in artificial intelligence. Datasets with too high imbalance ratio will cause the prediction performance degradation of the classification model. Many oversampling methods like SMOTE are proposed to solve the problem of sample imbalance. However due to the impact of noise samples and within imbalance distribution in datasets, performance of traditional oversampling methods were often limited. In this paper, we design a safety adaptive oversampling method combined with radial-based undersampling. Firstly, propose a region-partitioning criterion for the dataset, dividing it into safe region, boundary region, and noise region. Secondly, the undersampling strategy based on Gaussian radial basis function is combined with oversampling methods to remove noise samples and redundant majority samples that affect classification, and the number of redundant samples to be removed is determined based on the Pauta criterion. Afterwards, conventional synthetic samples are generated in the boundary region by a random linear interpolation strategy, while low variance synthetic samples are generated in the safe region by an affine shadowsampling strategy to reduce the offset of synthetic samples. Finally, in order to avoid the generation of noise samples, the deviation degree between each synthetic sample and their parent sample is discriminated by the concept of class potential, thereby ensuring a safer sample generation process. In experiments, 18 classic UCI imbalanced datasets are selected to compare proposed method with other oversampling methods. Experimental results show that our method has better performance than other listed oversampling methods.