一种基于带核随机子空间的聚类集成算法

发布时间：2018-04-01 01:11

本文选题：子空间聚类　切入点：聚类集成　出处：《南京大学学报(自然科学)》2017年06期

【摘要】：随机子空间聚类集成通过属性随机采样产生属性子空间,并将子空间上的基聚类结果进行集成得到最终聚类结果.在这一过程中,子空间产生的随机性虽然为聚类集成提供了很大的差异度,但是无法保证基聚类结果的有效性,这是因为随机产生的子空间有可能只包含极少的重要属性.针对这一不足,提出了一种带核随机子空间生成策略:首先依据粗糙集理论中的互补互信息选出对于刻画数据集整体信息至关重要的属性子集,作为每个属性子空间的"核心",再从剩余属性集中随机选择一定数量的属性与核心属性共同构成聚类子空间.这种策略在兼顾子空间之间差异性的同时也提高了每个属性子空间对数据整体信息的刻画能力,从而得到更好的聚类集成结果.在大量UCI数据集上的实验证实了所提方法的合理性和有效性.
[Abstract]:Random subspace clustering ensemble generates attribute subspace by attribute random sampling, and the result of base clustering on subspace is integrated to obtain the final clustering result.In this process, the randomness generated by subspace can not guarantee the validity of base clustering results, although it provides a great degree of difference for clustering integration, because randomly generated subspaces may contain only a few important attributes.To solve this problem, a strategy of generating random subspace with kernels is proposed. Firstly, based on complementary mutual information in rough set theory, attribute subsets are selected which are important for describing the global information of data sets.As the "core" of each attribute subspace, a certain number of attributes and core attributes are randomly selected from the residual attribute set to form a cluster subspace.This strategy not only takes into account the differences between subspaces, but also improves the ability of each attribute subspace to depict the whole information of the data, so as to obtain a better result of clustering integration.Experiments on a large number of UCI datasets show that the proposed method is reasonable and effective.
【作者单位】：山西大学计算机与信息技术学院;计算智能与中文信息处理教育部重点实验室;
【基金】：国家自然科学基金(61303008,61202018) 山西省自然科学基金(2013021018-1)
【分类号】：TP311.13

【相似文献】