基于多粒度粗糙集的聚类融合方法

现有的聚类融合算法从聚类成员的角度出发，若使用全部聚类成员则融合结果受劣质成员影响，对聚类成员进行选择再进行融合则选择的策略存在主观性。为在一定程度上避免这两种局限性，可以从元素的角度出发，提出一种新的聚类融合方法。通过多粒度决策不一致粗糙集来选择一部分类别确定的元素，再利用这部分元素进行聚类融合生成新的划分；多粒度决策不一致粗糙集模型能够刻画多粒度决策过程中属性一致而决策不一致的现象，提出了一种基于多粒度决策不一致的粗糙集模型，并给出了一种聚类融合方法。具体做法是：首先在数据集上多次使用K-means聚类算法，生成论域上的多个粒结构；其次对所有粒结构两两之间求粒间包含度，建立包含度矩阵，对矩阵使用Otsu算法计算阈值，得出多组满足阈值条件的信息粒，求解多粒度决策不一致下近似和上近似；最后分别处理下近似与边界域中元素的类别，从而获得了一个经过融合的聚类划分。实验结果表明，该方法能够有效改善聚类的结果，具有较高的时间效率，且算法具有较好的鲁棒性。

Clustering ensemble algorithm based on multi-granulation rough set

Existing clustering ensemble algorithm starts from the perspective of cluster members, if all the cluster members are used, the ensemble result is affected by the inferior members. If the cluster members are selected and then used in ensemble, the selected strategy has subjectivity. To avoid these two limitations to some extent, from the perspective of elements, this paper proposed a new clustering fusion method. It selectied a part of class-determined elements through multi-granulation rough sets with incongruous decisions, and then used this part of the elements to generate a new clustering. Multi-granulation rough set model with incongruous decisions could describe the phenomenon of inconsistent decisions with consistent attribute set. Therefore, this paper proposed a model of multi-granulation rough set with incongruous decisions and a clustering ensemble algorithm based on the model. First of all, it ran a K-Means clustering algorithm several times on the data set in the casennd, generated multiple granule structures: Next, it calculated inclusion degrees among all the granulations, and obtained the matrix of inclusion degree. Used Otsu''s method to generate a threshold, then got several group of granulation that met the threshold condition. According to the model of multi-granulation rough set with incongruous decision, it obtained lower and upper approximations. Finally, classified the elements of lower approximation and boundary separately to obtained a clustering that has been fused. The experiments showed that the algorithm had a high time efficiency and robustness, which improved the result of K-means clustering.