基于聚类欠采样的集成不均衡数据分类算法

发布时间：2018-03-25 07:01

本文选题：不均衡数据　切入点：欠采样　出处：《工程科学学报》2017年08期

【摘要】：传统的分类算法大多假设数据集是均衡的,追求整体的分类精度.而实际数据集经常是不均衡的,因此传统的分类算法在处理实际数据集时容易导致少数类样本有较高的分类错误率.现有针对不均衡数据集改进的分类方法主要有两类:一类是进行数据层面的改进,用过采样或欠采样的方法增加少数类数据或减少多数类数据;另一个是进行算法层面的改进.本文在原有的基于聚类的欠采样方法和集成学习方法的基础上,采用两种方法相结合的思想,对不均衡数据进行分类.即先在数据处理阶段采用基于聚类的欠采样方法形成均衡数据集,然后用AdaBoost集成算法对新的数据集进行分类训练,并在算法集成过程中引用权重来区分少数类数据和多数类数据对计算集成学习错误率的贡献,进而使算法更关注少数数据类,提高少数类数据的分类精度.
[Abstract]:Most of the traditional classification algorithms assume that the data set is balanced and pursue the overall classification accuracy, but the actual data set is often unbalanced. Therefore, the traditional classification algorithm is easy to cause a high classification error rate of a few kinds of samples when dealing with actual data sets. There are two kinds of improved classification methods for unbalanced data sets: one is to improve the data level. The method of over-sampling or under-sampling is used to increase or reduce the number of classes of data, the other is to improve the algorithm level. Using the idea of combining two methods, the unbalanced data is classified, that is, in the stage of data processing, the method of under-sampling based on clustering is used to form the balanced data set, and then the new data set is classified and trained by AdaBoost integration algorithm. In the process of algorithm integration, the weight is used to distinguish the contribution of a few class data from the majority class data to calculate the learning error rate of integration, which makes the algorithm pay more attention to a few data classes and improve the classification accuracy of a few class data.
【作者单位】：北京科技大学东凌经济管理学院;
【基金】：国家自然科学基金资助项目(71271027) 高等学校博士学科点专项科研基金资助项目(20120006110037)
【分类号】：TP311.13

【相似文献】