基于旋转平衡森林的不平衡数据分类算法

Classification algorithm of imbalanced data based on rotation balanced forest

针对不平衡数据中的分类问题,提出一种基于旋转森林的改进模型——旋转平衡森林(rotation balanced forest,ROBF)。以集成思想为核心,从数据层和算法层相结合的角度出发,针对Safe-Level-Smote方法中存在的模糊类边界问题采取两点改进:安全等级再划分机制;引入约束度不同的控制因子,经改进后得到Hyper-Safe-Level-Smote,将Hyper-Safe-Level-Smote与旋转森林模型相结合得到旋转平衡森林。通过在UCI的6组数据集上将5种算法进行对比,对比结果表明,ROBF算法在保持良好分类准确率的同时,有着更具竞争力的TPR和G-mean。该结果验证了ROBF算法在处理不平衡问题上的有效性。

To solve the classification problem in imbalanced data,an improved model based on rotation forest,namely rotation balanced forest(ROBF),was proposed.Based on the idea of ensemble,and from the perspective of combining the data layer and the algorithm layer,two improvements were designed to solve the fuzzy class boundary problem in the Safe-Level-Smote method including the security level re-division mechanism,and the introduction of control factors with different degrees of constraint.After the improvement,the Hyper-Safe-Level-Smote was obtained.The Hyper-Safe-Level-Smote was combined with the rotation forest model to obtain a rotation balanced forest.Six data sets were selected from the UCI,and the five algorithms were compared to each other.The results show that the ROBF algorithm maintains good classification accuracy with more competitive TPR and G-mean.This result verifies the effectiveness of the ROBF algorithm in dealing with imbalance problems.