Abstract:

Aiming at discovering complementarity and consistency among multi-view data, a multi-view clustering based on Graph Convolutional neural Network（GCN） is proposed. By adding pairwise constraints on embeddings learned via the GCN on common sub-graphs of multiple views, consistency can be effectively measured. Through sharing the parameters of GCN, generating embeddings of each view based on their complete graphs, and concatenating those embeddings for multi-view embedding, complementarity will be explored. Besides, a Kullback-Leibler（KL） divergence based objective is designed to constrain the above embedding, leading to clustering oriented embedding learned. Experiments are conducted on five widely used datasets, achieving best clustering results, which clarifies the effectiveness of the method for exploring complementarity and consistency among multi-view data.

Key words: multi-view clustering, graph convolutional neural network, Kullback-Leibler divergence

摘要：

针对多视角数据间互补与一致特性难以刻画问题，提出一种基于图卷积神经网络的多视角聚类方法。通过对样本不同视角间相同邻接子图基于图卷积神经网络学习到的表达进行约束，有效挖掘了多视角数据间的一致特性。通过共享图卷积神经网络参数、学习不同视角完整邻接图嵌入表达并串接得到多视角表达，有效挖掘了多视角数据间的互补特性。对上述多视角表达增加相对熵约束，使得最终学习到的多视角表达得以提升并符合聚类特性。在五个数据集上均取得了最好的聚类效果，说明所提出的基于图卷积神经网络的聚类方法可以有效挖掘视角间互补与一致特性并提升聚类性能。

关键词: 多视角聚类, 图卷积神经网络, 相对熵