基于Hellinger距离的混合数据集中分类变量相似度分析

分类变量的相似度分析是数据挖掘任务中的一个重要环节,现有的分类变量相似度算法中存在忽视变量差异、受不均衡分布影响严重、无法应用于混合数据集等缺点。为克服以上缺点,提出了一种基于Hellinger距离的分类变量相似度算法。该算法累加分类变量对应子集中不同属性变量的分布差异作为相似度,且支持混合数据集。将所提算法代入聚类算法并应用于UCI公共数据集,结果表明,该算法在准确度、有效性和稳定性上都有较大提高。

Hellinger Distance Based Similarity Analysis for Categorical Variables in Mixture Dataset

Similarity analysis of categorical variables is an important part of data mining.The traditional methods have the defects of neglecting the difference between categorical variables,which are seriously affected by unbalanced dataset and can not be used in mixture dataset.To overcome the shortcomings mentioned above,this paper proposed an algorithm to measure the similarity between categorical variables based on the Hellinger distance.It accumulates the distribution differences of variables with different attributes in subsets corresponding to categorical variables as similarity variables and fits for mixture dataset.The experiments which use the derived similarity metrics in clustering algorithm and apply UCI datasets show that there is significant improvement in accuracy,validity and stability.