一种基于多属性权重的分类数据子空间聚类算法

采用多属性频率权重以及多目标簇集质量聚类准则，提出一种分类数据子空间聚类算法.该算法利用粗糙集理论中的等价类，定义了一种多属性权重计算方法，有效地提高了属性的聚类区分能力；在多目标簇集质量函数的基础上，采用层次凝聚策略，迭代合并子簇，有效地度量了各类尺度的聚类簇；利用区间离散度，解决了使用阈值删除噪音点所带来的参数问题；利用属性对簇的依附程度，确定了聚类簇的属性相关子空间，提高了聚类簇的可理解性.最后，采用人工合成、UCI和恒星光谱数据集，实验验证了该聚类算法的可行性和有效性.

A Subspace Clustering Algorithm of Categorical Data Using Multiple Attribute Weights

In this paper, we propose a subspace clustering algorithm using frequencies of multiple attributes of categorical data. An attribute-value weight is calculated on the basis of equivalence class in rough set theory by using frequencies of multiple attributes. Attribute-value weights offer ample opportunities to improve classiffication ability of clustering. The well-known parameter problem, which is caused by using a threshold to delete noise points, is solved by the virtue of interval dispersion degrees. By adopting the hierarchical clustering method to iteratively merge sub-clusters, we effectively measure various scale clusters on the basis of a multi-objective clusters quality function. An attribute subspace is determined with the relevance degree of dimension to a cluster, so as to improve the cluster's interpretability. Finally, we validate the feasibility and effectiveness of our algorithm through extensive experiments using synthetic data as well as UCI and stellar spectral data sets.