基于商空间理论的非平衡数据集分类算法

Classification algorithm for imbalance dataset based on quotient space theory

在机器学习及其分类问题时经常会遇到非平衡数据集,为了提高非平衡数据集分类的有效性,提出了基于商空间理论的过采样分类算法,即QMSVM算法。对训练集中多数类样本进行聚类结构划分,所得划分结果和少数类样本合并进行线性支持向量机(SVM)学习,从而获取多数类样本的支持向量和错分的样本粒;另一方面,获取少数类样本的支持向量和错分的样本,进行SMOTE采样,最后把上述得到的两类样本合并进行SVM学习,这样来实现学习数据集的再平衡处理,从而得到更加合理的分类超平面。实验结果表明,和其他几种算法相比,所提算法虽在正确分类率上有所降低,但较大改善了g_means值和acc+值,且对非平衡率较大的数据集效果会更好。

The application of data classification is usually confronted with a problem named imbalanced dataset in the machine learning.To improve the performance of imbalanced dataset classification,the over-sampling classification algorithm based on quotient space theory（QMSVM） was proposed.The algorithm partitioned majority data on clustering structure,and combined the results and minority data for linear Support Vector Machine（SVM） learning.Support vectors and sample of fault of majority data were obtained from those granules.On the other hand,support vectors and sample of fault of minority data were obtained and the Synthetic Minority Over-sampling Technique（SMOTE） was adopted.Thus,two new kinds of samples were merged for SVM learning,so as to rebalance the training set and get a more reasonable classification of hyperplanes.The experimental results show that,in comparison with several other algorithms,the accuracy of the proposed algorithm decreases,but it significantly improves the g_means value and classification accuracy of positives and the effect is better on the imbalance rate of larger datasets.