基于一趟聚类的不平衡数据下抽样算法

抽样是处理不平衡数据集的一种常用方法,其主要思想是改变类别的分布,缩小稀有类与多数类的分布比例差距.提出一种基于一趟聚类的下抽样方法,根据聚类后簇的特征与数据倾斜程度确定抽样比例,按照每个簇的抽样比例对该簇进行抽样,密度大的簇少抽,密度小的簇多抽或全抽.在压缩数据集的同时,保证了少数类的数量.实验结果表明,本文提出的抽样方法使不平衡数据样本具有较高的代表性,聚类与分类性能得到了提高.

Under-sampling Method Based on One-pass Clustering for Imbalanced Data Distribution

Sampling is a widely used method in dealing imbalanced dataset.The main idea of sampling is changing the distribution of various classes in a dataset,decreasing the difference in the distribution between majority class and minority class.This paper proposes a under-sampling approaches based on one-pass clustering for imbalance problem.According to the features of each clusters and the degree of data inclination,a sample Ratio is made for each clusters.Hence each cluster can be sampled with its sample ratio.A Cluster with greater density will be sampled into comparatively less selection.In contrast,a Cluster with smaller density will be more or entirely sampled.With decreasing the amount of the whole dataset,the amount of minority class can be ensured.The experimental result show that our sample approaches can make the sample form the imbalanced data more typical and improve the performance in classification and clustering.