基于属性关系矩阵的AP子空间聚类算法

AP subspace clustering algorithm based on attributes relation matrix

仿射传播(Affinity Propagation,AP)聚类算法将所有数据点作为潜在的聚类中心,在相似度矩阵的基础上通过消息传递进行聚类,但不适用于子空间聚类。基于属性关系矩阵的AP子空间聚类算法(ARMAP)是一种异步软子空间聚类算法,通过计算属性α的α-β邻域得到属性的关系矩阵,查找极大全1子矩阵得到数据集的兴趣度子空间,在各兴趣度子空间使用AP算法聚类,完成子空间聚类的任务。ARMAP算法将子空间的查找转换成查找矩阵的极大全1子矩阵,在正确查找子空间的同时降低了时间复杂度,既保留了AP聚类算法的优点,又克服了AP算法不能进行子空间聚类的不足。

Affinity propagation （AP） algorithm takes all data as potential clustering centers. Clustering is carried out by message passing based on the similarity matrix. But it is not appropriate for subspace clustering. AP subspace clustering algorithm based on attributes relation matrix （ARMAP） is an asynchronous soft subspace clustering algorithm. This algorithm calculates attribute relation matrix through α-β neighborhood of attribute a. The candidate of all interesting subspaces is achieved by looking for the maximum sub-matrixes of attribute relation matrix which contain only 1. All subspace clusters can be obtained through AP clustering on interesting subspaces. The method obtains interesting subspaces correctly and reduces time complexity at the same time. It not only keeps the advantages of AP clustering algorithm, but also overcomes the shortcomings of AP algorithm which can not be used for subspace clustering.