一种针对不平衡数据分类的集成学习算法

An ensemble learning algorithm for unbalanced data classification

针对水声目标识别中常被忽略的数据不平衡问题,提出一种随机子空间AdaBoost算法(RSBoost)。通过随机子空间法在不同水声特征空间上提取子训练样本集,在各个子训练样本集上训练基分类器,将其中少类间隔均值最大的基分类器作为本轮选定的分类器,迭代形成最终集成分类器。在实测数据上进行试验,利用F-measure和G-mean两个准则对RSBoost算法和AdaBoost算法在不同特征集上的性能进行评价。试验结果表明:相对于AdaBoost算法,RSBoost算法在F-measure准则下的平均值由0.07提升到0.22,在G-mean准则下的平均值由0.18提升到0.25,说明在处理水声数据不平衡分类问题上,RSBoost算法优于AdaBoost算法。

For unbalanced data classification problem in underwater acoustic target recognition, a random subspace AdaBoost algorithm called RSBoost was proposed. Subtraining sample set was extracted by random subspace method in different underwater acoustic feature space and base classifier was trained in every subtraining sample set. The base classifier with the maximum margin mean of minority class was chosen as the base classifier of this round, the final ensemble classifier was formed iteratively. The experiment was carried out on the measured data, the performance of RSBoost and AdaBoost in different feature space was evaluated by F-measure and G-mean. The results showed that, compared with AdaBoost, the F-measure of RSBoost improved from 0.07 to 0.22 and the G-mean improved from 0.18 to 0.25, which showed that RSBoost was superior to AdaBoost in underwater acoustic unbalanced classification problem.