一种邻域粒谱聚类方法

A Neighborhood Granular Spectral Clustering Method

谱聚类是一种无监督学习的聚类方法,其具有能够收敛至全局最优且适用于任意形状样本空间的优点.然而,传统方法构造的相似矩阵有时难以准确反映出数据之间的近似关系,从而导致聚类结果不佳.粒计算技术能够很好地解决这一问题.通过将数据邻域粒化,从粒子的视角重新衡量数据之间的近似关系,提出了一种基于邻域粒的谱聚类方法.首先,将样本的单一属性通过邻域粒化的方式形成邻域粒子;然后,将属于同一样本的粒子组合构造成粒子向量;接着,利用定义的2种邻域粒距离公式,对构造出的粒向量进行距离度量,并通过径向基函数生成相似矩阵,从而进行谱聚类;最后,使用UCI数据集进行验证,将谱聚类算法与邻域粒结合,从邻域参数和邻域粒向量的距离度量方式2个方面进行性能测试,并与传统聚类算法进行对比.实验结果表明,基于邻域粒构造的相似矩阵在谱聚类中是可行且有效的.

Spectral clustering is an unsupervised learning clustering method,which has the advantages of convergence to the global optimum and is applicable to arbitrary shape sample space.However,the similarity matrix constructed by traditional methods can not reflect the approximate relation between data sometimes,so the clustering results are not good.Granular computing can solve this problem well.By granulating the data in the neighborhood and re-measuring the approximate relation between the data from the perspective of granules,a spectral clustering method based on neighborhood granules is proposed.Firstly,the single attribute of the sample was formed into the neighborhood granules by the way of neighborhood granulation,and then the granule vector was constructed by combining the granules belonging to the same sample.By using two kinds of neighborhood granule distance formula defined,the constructed grain vector was measured by distance,and the radial basis function was used to generate a similar matrix for spectral clustering.Finally,the performance of the spectral clustering algorithm combined with neighborhood granules was tested using the UCI datasets for validation.The algorithm's performance was evaluated in two aspects:neighborhood parameters and distance measurement methods of neighborhood granule vectors.The results were compared with those of traditional clustering algorithms.The experimental results showed that the similarity matrix constructed using neighborhood granulation is feasible and effective for spectral clustering.