【摘要】 多视图聚类(Multi-View Clustering,MVC)旨在利用不同视图间的一致性和互补性来高效处理多视图数据,是大数据分析中重要的研究方向之一.然而,现有方法无法有效学习到多视图信息间的潜在联系,且缺乏考虑视图重要性差异问题.针对上述这些问题,本文提出了一种基于分布对齐变分自编码器的深度多视图聚类方法(Deep Multi-View Clustering based on Distribution Aligned Variational Autoencoder,DMVCDA).首先,针对特定视图我们利用多个变分自编码器从不同视图中提取潜在特征,并对特征的分布进行对齐,以挖掘包含基本信息的潜在特征;然后,引入视图权重参数,获取共享的潜在特征;最后,在潜在特征上建立面向聚类的损失目标,使得学习到的潜在特征更适合聚类任务,从而提高聚类精度.在五个公共多视图数据集上的实验结果表明,我们的模型在精确度(ACC)、标准互信息(NMI)和纯度(Purity)等多个聚类评价指标上均表现出优异的性能.更多还原

【Abstract】 Multi-view data means that the same object can be described by multiple different data sources or features,and each data source or feature can be viewed as a specific view.Multi-view clustering aims to exploit the consistency and complementary information among different views to efficiently process multi-view data,which is one of the most important research topics in big data analysis.Recently,a surge of multi-view clustering methods have been developed and achieved promising success, which has attracted considerable attentions in machine learning and data mining community.However,most of the existing methods neither effectively learn the latent relationship among multiple views nor consider the different importance of each view.By such,the solution to multi-view clustering is often sub-optimal.In order to address these problems,this paper proposes a deep multi-view clustering based on distribution aligned variational autoencoder so as to improve performance of multi-view clustering.First,we use view-specific variational autoencoder to extract latent features from different views,and align the learned view data distribution to further mine latent features containing basic information.Then,we introduce view weight parameters to fuse the view-specific features into a shared latent one.Finally,the loss function for clustering is established on the latent features,so that the learned latent features are more suitable for clustering tasks,thereby improving the clustering accuracy.The experimental results on five common multi-view datasets show that our model has achieved excellent performance in terms of multiple clustering evaluation metrics,such as accuracy(ACC),normalized mutual information(NMI) and purity(Purity),which validates the effectiveness of our model.更多还原