基于特征相关度分析的不平衡数据混合采样方法

ata information balance through sampling is a major way to solve the problem of data imbalance. Based on the idea of mixed sampling, a new method of mixed sampling for imbalanced datasets based on feature correlation analysis is proposed in this paper. First, set the ratio of the important features of the dataset to all features, by comparing the correlation coefficients of each feature attribute with the sample category, a feature subset with a large correlation coefficient is selected as an important feature set according to the ratio coefficient. Then, take the median of all kinds sample data volume as the demarcation point, and mark the sampling markers of each sample, and combined with the selected important feature subset, the sampling strategies of over-sampling and under-sampling are determined respectively. Finally, the sampling balance coefficient is set, and the difference between the data of each sample before sampling and the median of the sample size is used to determine the data of each sample after sampling. On two sets of KEEL common data sets, the data sets without sampling, using conventional mixed sampling method and using the mixed sampling method mentioned are compared with three commonly used classification algorithms with distinct strategy differences in this paper. The experimental results show that the classification accuracy of the data set processed by the proposed method is much better than that of the other two groups of experiments under the three classification algorithms, which verifies the effectiveness and reliability of the proposed method.