Self-Supervised Graph Convolutional Network for Multi-View Clustering
Self-Supervised Graph Convolutional Network for Multi-View Clustering
复制标题
DOI：
10.1109/tmm.2021.3094296
复制
发表时间：
2021-07
影响因子：
7.3
通讯作者：
Wei Xia;Qianqian Wang;Quanxue Gao;Xiangdong Zhang;Xinbo Gao
中科院分区：
文献类型：
--
作者：
Wei Xia;Qianqian Wang;Quanxue Gao;Xiangdong Zhang;Xinbo Gao
现有的基于图卷积网络（GCN）的多视图学习方法直接使用图结构作为视图描述符，这可能会抑制多媒体数据的多视图学习能力。主要原因是，在实际的多媒体应用中，图形结构可能包含异常值。此外，他们没有利用所提出的方法获得的不准确聚类标签中嵌入的信息，导致聚类结果较差。这些观察结果促使我们研究是否有更好的基于GCN的多视图聚类框架。为此，本文提出了一种端到端自监督图卷积多视图聚类网络（SGCMC）。具体而言，SGCMC通过欧拉变换将原始节点内容映射到复空间，构建了一种新的图结构数据视图描述符，不仅可以抑制异常值，还可以揭示数据中嵌入的非线性模式。同时，本文提出的SGCMC利用聚类标签指导潜表示和系数矩阵的学习，潜表示和系数矩阵用于后续节点聚类。通过这种方式，聚类和表示学习无缝连接，目的是获得更好的聚类结果。大量的实验结果表明，所提出的SGCMC优于目前最先进的方法。
Despite the promising preliminary results, existing graph convolutional network (GCN) based multi-view learning methods directly use the graph structure as view descriptor, which may inhibit the ability of multi-view learning for multimedia data. The major reason is that, in real multimedia applications, the graph structure may contain outliers. Moreover, they fail to take advantage of the information embedded in the inaccurate clustering labels obtained from their proposed methods, resulting in inferior clustering results. These observations motivate us to study whether there is a better alternative GCN based framework for multi-view clustering. To this end, in this paper, we propose an end-to-end self-supervised graph convolutional network for multi-view clustering (SGCMC). Specifically, SGCMC constructs a new view descriptor for graph-structured data by mapping the raw node content into the complex space via Euler transformation, which not only suppresses outliers but also reveals non-linear patterns embedded in data. Meanwhile, the proposed SGCMC uses the clustering labels to guide the learning of the latent representation and coefficient matrix, and the latter in turn is used to conduct the subsequent node clustering. By this way, clustering and representation learning are seamlessly connected, with the aim to achieve better clustering results. Extensive experimental results indicate that the proposed SGCMC outperforms the state-of-the-art methods.