【作者】 王勇；

【导师】 古平；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2017， 硕士

【摘要】 数据分类作为数据挖掘中备受关注的一个研究方向,已经有了许多研究成果,这些成果大多在数据集中样本分布均衡的条件下能够取得很好的结果,但是在实际场景中,常见的需要被分类的数据集在分布状态上普遍都有不均衡的特点。所谓不均衡,指的是在整个数据集中,某个类别占的比例相对比较大,其他类别占的比例相对较小,占比大的样本实例一般划分为多数类,占比小的样本实例则划分为少数类。在类别比例不均衡的数据集里,少数类样本相对决策边界的分布具有差异性,越靠近决策边界被错分的可能性越大,基于此,本文提出通过数据集构造维诺图,按照少数类样本相对于决策边界的分布差异,赋予每个少数类样本不同的权重,计算权重符合规则的样本的采样概率,随机选取样本合成人工少数类。基于维诺图的不均衡数据集分类主要提出以下改进:1.新的边界识别方法。不均衡数据集中越靠近决策边界的少数类样本分类意义越高,传统的分类算法并未对这种差异性做出比较妥善的处理。本文通过构造维诺图找到分隔少数类与多数类样本之间的维诺边作为近似决策边界集合,计算每个少数类样本到边界集的最小距离作为少数类样本的边界度;2.基于边界度的采样策略。根据边界样本集确定新的边界,利用新的边界对边界度进行一定的变换后代入以自然常数e为底的指数函数,利用所有样本的函数值进行归一化,最终得到每个少数类样本的采样概率,然后随机选择样本进行过采样。上述两步称之为V-synthⅠ算法。3.处理局部不均衡。上述算法使用样本到决策边界的距离作为权重来划分边界样本,这种方式更灵活、准确。但是只根据少数类样本的分布差异计算采样概率,没有考虑多数类分布对少数类的影响,可能出现数据集整体均衡而局部不均衡的现象。为此,在V-synthⅠ算法的基础上使用层次聚类中的凝聚类方法对多数类聚类,形成若干个簇,计算每个簇中多数类的分布密度以及多数类簇对少数类样本点的影响因子,更新样本的采样概率。该算法称之为V-synthⅡ算法。通过人为构造特殊分布的数据集和选取分类问题中常用的UCI数据集用于实验分析,利用上述的两种算法分析各类别占比不均衡的数据集可以取得比较理想的分类结果。更多还原

【Abstract】 Classification is a very important research direction in data mining which had many research achievements,in the condition of data sample distribution equilibrium conditions these achievements can achieve good results,but in the actual scene,most of the data sets are imbalanced.Imbalanced data set mean include one excellent much catatgory samples,a small number of other categories samples,the former is called the majority class,the latter is called the minority class.When dealing with imbalanced data sets,the existing classification algorithms can not accurately identify the category of each sample.In the imbalanced data set,the distribution of minority class sample relative to decision boundary is different,closer to the possibility of wrong decision boundary is bigger,based on this,this paper proposes an novel algorithms with Voronoi diagram,according to the distribution of minority class sample relative to the decision boundary,giving each minority class samples with different weights,calculate the samples’ s weight that fit an rule,then randomly selected sample to synthetic minority class.The classification of imbalanced data set based on Voronoi diagram,mainly include the following improvements:1 A new method for boundary recognition.The higher the classification of imbalanced data sets closer to the decision boundary,traditional classifications algorithm does not deal with the difference property.This paper sets to find approximate decision boundary which separate minority class and majority class samples by constructing the Voronoi diagram with data set,the minimum distance is calculated for each minority class samples as its boundary degree;2 Sampling strategy based on boundary degree.According to the boundary samples to determine the new boundary,transform boundary degree to exponential function with natural constant and normalized the function values,get the sampling probability of each minority class sample,then the sample was randomly selected for sampling.The two step is called V-synthⅠ.3 Deal with local imbalance.The algorithm uses the distance from the sample to the decision boundary as the weights to divide the boundary samples,which is more flexible and accurate,and can greatly reduce the possibility of the synthetic noise samples.However,only considering the sampling probability of a small number of samples,without considering the characteristics of the majority class distribution,the overall equilibrium of the data set and the phenomenon of local imbalance may occur.Therefore,by clusting majority class samples,form some clusters,calulate the influence factor that the distribution density of each cluster to minority class sample,update the probability sample.This algorithm is called V-synthⅡ.Through construct special distribution of the data set and the data set which commonly used in UCI for the experimental analysis,the use of the above two algorithms to deal with the data set can get better classification results.更多还原