| 摘要: |
| 多视角子空间聚类方法为高维多视角数据的聚类问题提供了大量的解决方案. 但是现有的子空间方法仍不能很好地解决以下两个问题: (1) 如何利用不同视角的差异性进行学习获得一个优质的共享系数矩阵; (2) 如何增强共享系数矩阵的低秩性. 针对以上问题, 提出了一种有效的双加权多视角子空间聚类算法. 该算法首先通过子空间自表达学习到每个视角的系数矩阵, 然后采用自适应权重策略构建一个共享系数矩阵, 最后利用加权核范数逼近系数矩阵的秩, 使得系数矩阵的表示更加低秩, 进而取得更好的聚类结果. 采用增广拉格朗日乘子法来优化目标函数, 并在6个广泛使用的数据集上进行实验, 验证了该算法的优越性. |
| 关键词: 多视角子空间聚类 系数矩阵 权重 加权核范数 低秩 |
| DOI：10.13328/j.cnki.jos.006148 |
| 分类号:TP18 |
| 基金项目:国家自然科学基金(61573273) |
|
| Dual Weighted Multi-view Subspace Clustering |
|
CAO Rong-Wei1, ZHU Ji-Hua1, HAO Wen-Yu1, ZHANG Chang-Qing2, ZHANG Zhuo-Han1, LI Zhong-Yu1
|
|
1.School of Software Engineering, Xi'an Jiaotong University, Xi'an 710049, China;2.College of Intelligence and Computing, Tianjin University, Tianjin 300350, China
|
| Abstract: |
| In order to solve the problem of clustering multi-view data, many multi-view subspace clustering methods have been proposed and achieved great success. However, they cannot well fix the following two problems. 1) How to leverage the difference between different views to learn a shared coefficient matrix with high quality. 2) How to further enforce the low rank property of the common coefficient matrix. To handle the above problems, an effective method dubbed dual weighted multi-view subspace clustering is proposed. In detail, the coefficient matricesare first learned for each view by self-representation model, and then they are fusedinto a common representation with a self-weighted strategy, finally weighted nuclear norm instead of nuclear norm is employed to approximate the rank of the common coefficient matrix, so that the performance of clustering can be improved. An augmented Lagrange multiplier based optimal algorithm is imposed to solve the established objective function. Experiments conducted on six real world datasets validate the superiority of the proposed method. |
| Key words: multi-view subspace clustering coefficient matrix weighted weighted nuclear norm low rank |
