多视角表示学习是针对多视角数据集的特征提取方法，旨在针对同一对象具有多个不同描述视角的数据集进行特征提取。它能够探索视角之间的联系，整合利用不同视角下的信息，并带来对不同领域问题的深入理解。与此同时，多视角数据的异质性也给研究带来了一些独特的挑战。

虽然目前关于多视角表示学习的研究成果颇丰，但现有算法通常侧重于分类或聚类任务。不同的是，本文专注于无监督的表示学习，旨在提出一个具有普适性的特征提取网络，使得到的多视角特征能够在不同的后续任务中都具备良好的应用效果。

在无监督的多视角特征提取过程中，许多学者认为不同视角之间的具有指导意义的特征存在一定联系，并且这种联系对问题的解决具有指导意义，从而发展成基于相关性的多视角表示学习这一分支。然而传统的相关性模型存在一些不足，比如只能度量视角间变量的线性关系，并且忽略了视角间冗余信息，重复计算视角间的共性信息。

为解决以上问题，本文提出一个全新的无监督特征提取方法，改进多视角数据的特征表示，称为深度Copula典型相关的多视角表示学习框架。该框架有以下两项创新之处。

（1）构建了基于Copula熵的深度相关性分析与特征提取网络。

该网络通过改进传统特征提取网络中的正则化项，增强特征提取网络的非线性性。首先，引入Copula理论来构造视角间非线性相关矩阵，提出基于Copula熵的典型相关分析方法，优化视角间非线性相关性的度量。其次，将最大化视角间非线性相关性作为网络的优化目标，从而构建深度Copula典型相关的多视角特征提取网络。

（2）提出了基于重构的多视角特征融合网络。

为消除视角间信息冗余，降低多视角特征向量的维度，建立了多视角的特征重构网络。利用自编码器，将多视角特征提取网络的输出特征编码成一个完整的多视角特征表示向量。

以上多视角特征提取网络与特征重构网络就构成了深度Copula典型相关的多视角表示学习网络。该框架考虑了视角间的非线性相关性，并且能消除视角间信息冗余。所得到的多视角特征能够结合对于特定视角的编码和多视角的共同特征编码，平衡视角之间的互补性和一致性。

Multi-view representation learning is a feature extraction method for multi-view datasets, aiming at extracting features from datasets with multiple different views from the same object. It can explore the connections between views, integrate and utilize information from different views, and bring about in-depth understanding of phenomena in different fields. At the same time, the heterogeneity of multi-view data also brings some unique challenges to the research.

Although there are a lot of research results on multi- view representation learning, the existing algorithms usually focus on classification or clustering tasks. The difference is that this paper focuses on unsupervised representation learning, aiming to propose a universal feature extraction network, so that the obtained multi-view features can have good application effects in different follow-up tasks.

In the process of unsupervised multi-view feature extraction, many scholars believe that there is a certain relationship between the guiding features of different views, and this relationship is of guiding significance to the solution of the problem, thus developing the branch of multi-view representation learning based on correlation. However, traditional correlation models have some shortcomings, such as only measuring the linear relationship of variables between views, ignoring the redundant information, and repeatedly calculating the common information between views.

To solve the above problems, this paper proposes a new unsupervised feature extraction method to improve the feature representation of multi-view data, which is called multi-view representation learning with deep Copula canonical correlation analysis. This framework has two innovations.

(1) Constructing a deep correlation feature extraction network based on Copula entropy.

By improving the regularization term in the traditional feature extraction network, the measure of nonlinear correlation between views is optimized to enhance the nonlinearity of the feature extraction network. Firstly, the Copula theory is introduced to construct the nonlinear correlation matrix between views. Then, with the help of the canonical correlation analysis, the matrix is used to calculate the multivariate nonlinear correlation between views, and a canonical correlation analysis method based on Copula entropy is proposed. Secondly, by maximizing the nonlinear correlation between views as the optimization goal of the network, a multi-view feature representation network with deep Copula canonical correlation is constructed.

(2) Proposing a multi-view feature fusion network based on reconstruction.

In order to eliminate the information redundancy between views and reduce the dimension of multi-view feature vectors, a multi-view feature reconstruction network is established. An autoencoder is used to encode the output features of the multi-view feature extraction network into a complete multi-view feature representation vector.

The above multi-view feature extraction network and feature reconstruction network constitute the multi-view representation learning network, called multi-view representation learning with Copula canonical correlation analysis. The framework takes into account the nonlinear correlation and eliminates the information redundancy between views. The obtained multi-view feature is combined with the coding of specific views and the common feature coding of multiple views, balancing the complementarity and consistency between views.