基于自表征和群组效应的子空间聚类算法

Self-representation and grouping effect for subspace clustering

为解决目前聚类算法对噪声敏感和缺乏考虑样本间相关性等问题,提出一种充分考虑样本间相关性,使构造的关联矩阵保持子空间结构的子空间聚类算法。利用2,1-norm对每个样本进行自表征;群组效应确保相近样本的自表征系数亦相近,生成块对角化的样本自表征系数矩阵;根据自表征矩阵得到关联矩阵,在谱聚类模型下实现数据聚类。在Hopkins155等数据集上的实验结果表明,在聚类错误率评判标准下,该算法优于现有经典子空间聚类算法SRC、SSC等。

To solve the issues that previous clustering methods are sensitive to noise and fail to consider the correlations among samples,a subspace clustering algorithm was proposed by taking the correlations among samples into account,so that the simila-rity matrix of the proposed clustering method preserved the structure of subspace.An/2,i^norm was utilized to represent each sample by training samples.The grouping effect of the data was designed to ensure that the coefficient of close samples was simi-lar,aiming at generating a diagonal block self-representation coefficient matrix.An affinity matrix was obtained for conducting spectral clustering.Experimental results on datasets such as Hopkinsl55indicate that the proposed algorithm significantly out-performs the state-of-the-art methods,such as SRC and SSC.