UNDERSAMPLING BOOSTING FRAMEWORK FOR UNBALANCED PROBLEMS BASED ON ENTROPY AND CONFIDENCE

In order to solve the problems of boundary overfitting, poor generalization performance and important information loss in traditional methods, an undersampling boosting framework for unbalanced problems based on entropy and confidence is proposed. The dynamic up-sampling method was integrated with boosting to solve the boundary over fitting problem and improve the generalization performance. The confidence and entropy were used as the benchmark to ensure the validity and structure distribution of most samples in the process of undersampling. In addition, the generalization ability of dynamic sampling method was improved by the proposed boosting framework based on confidence degree further. Experimental results on two large datasets show the effectiveness of the proposed method.