Research On Multi-view Clustering Algorithm Based On Representation Learning

With the vigorous development of information and communication technology,it is easy for humans to obtain massive unlabeled multi-view data when they perceive the surrounding environment from multiple information sources.However,even though the industrial reform of integrated artificial intelligence（AI）has achieved remarkable results,there is still a shortcoming,such as prominent data barriers.This problem boils down to one thing: the performance of downstream tasks relies heavily on many labeled training examples,such as those required for ChatGPT and other foundation models.The full potential of multi-view data cannot be realized in the absence of labeled examples,which has become a bottleneck problem limiting the implementation of AI scenarios.Thus,it is of great value for human society to study efficient and effective unlabeled multi-view data labeling.Multi-view clustering seeks to uncover the complementary,similarity,and associative information embedded in data from different views to advance clustering performance.Designing an effective multi-view representation learning method to address the ‘heterogeneity gap’ dilemma between different views is an essential research component for multi-view clustering tasks.Numerous multi-view clustering methods based on representation learning have been proposed and achieved promising results.However,lots of challenges remain,including how to efficiently cluster large-scale multi-view data,collaborate with data features and structured graphs for richer view-consensus representation,effectively exploit information hidden in clustering labels,and make multi-view clustering models transparent and interpretable.This dissertation aims to learn discriminative multi-view consistent representation and conduct systematic research from shallow and deep representation learning aspects to develop more general multi-view clustering theories and models to tackle these complex issues.The specific research contents are listed as follows:1)To address the shortcomings of graph-based multi-view clustering methods,which have high computational complexity and do not simultaneously utilize the similarity information embedded within and between views,we propose a technique called tensorized bipartite graph learning for multi-view clustering（TBGL）.Firstly,we design a variancebased de-correlation anchor selection strategy（VDA）to ensure that the selected anchors cover all classes effectively and represent the data’s inherent structure.Moreover,the tensor Schatten p norm penalty is introduced to explore the complementary information and spatial structure between different views.The （?）_1,2 norm penalty and connectivity constraint are combined to examine the similar information within each view.As a result,the learned view-consensus bipartite graph can better describe the cluster structure and have connected components.Lastly,an efficient algorithm is developed to optimize the proposed TBGL.Experimental results on ten datasets demonstrate that the clustering accuracy and efficiency of the proposed TBGL achieve optimal performance at that time and effectively alleviate the problem of low efficiency of tensorized graph learning methods in large-scale data analysis.2)To address the shortcomings of unsupervised deep multi-view clustering methods that only utilize data features or structured graphs for representation learning and overlook the discriminative information embedded in clustering labels,we propose a technique called self-supervised graph convolutional network for multi-view clustering（SGCMC）.Firstly,we adopt the Euler transform to augment the multi-view description for attributed graph data,which suppresses outliers and uncovers the nonlinear patterns embedded in the original data.Next,we construct a multi-view graph convolutional autoencoder network to extract view-specific representations by simultaneously encoding each view’s data features and structured graphs,fully exploring the internal relationship between data features and structured graphs.Following this,we fuse view-specific representations in latent space to obtain a consistent representation for clustering.To alleviate the difficulty of unsupervised graph convolutional network optimization,we design a dual self-supervised module that uses clustering labels to supervise the learning of view-consensus and view-specific representations.Numerous experimental results demonstrate that these strategies significantly improve multi-view clustering performance.Finally,to further enhance the discriminability of view-consensus representation,we introduce a block diagonal structure constraint to refine the approach.As a result,we propose a multi-view graph embedding clustering network: joint selfsupervision and block diagonal representation（MVGC）.Quantitative experiments and visualization verification results indicate that this strategy promotes multi-view representation learning capability and improves clustering performance.3)Considering the shortcomings of deep multi-view representation learning,i.e.,the underutilization of pseudo-labels and the inability to perform one-step clustering,we propose a self-consistent contrastive multi-view clustering framework with pseudo-label guidance.We initially designed a multi-view cluster structure maintenance contrastive learning module utilizing the pseudo-label guidance information.In other words,the data representation consistency within and between views must maintain self-consistency with the cluster structure reflected by the pseudo-label.This approach helps alleviate the issue of inadequate cluster structure in multi-view representation learning.Additionally,we construct a contrastive clustering network that maps the view-consensus representation directly to the clustering space,effectively avoiding post-processing and addressing the out-of-sample problem.Experimental results demonstrate that our framework performs optimally on four challenging multi-modality datasets.To further validate the framework’s effectiveness in handling multi-view attributed graph data,we propose a self-consistent contrastive attributed graph clustering network with pseudo-label guidance（SCAGC）.Employing a cluster space contrastive learning strategy enhances the model’s ability to cluster graph structure data.Clustering results on various graph structure datasets reveal that the proposed SCAGC outperforms the most competitive methods and exhibits strong practicality.4)To effectively address the challenging problem of interpretability in deep multi-view clustering methods,we introduce an explainable multi-view clustering network based on causal invariance constraint（X-MVC）.At the representation level,based on the Principle of Common Cause and Independent Causal Mechanisms（ICM）,it constructs a representation causal invariance constraint and the independent constraint of representation dimensions.This enables the learning of multi-view representations with causal factor characteristics,thereby enhancing the interpretability of multi-view representation learning.Furthermore,we establish an interpretable multi-view clustering mechanism at the decision-making level.Employing weights and biases decoupling strategy in the neural network,this mechanism differentiably reconstructs the classic multi-view K-Means.Simultaneously,we design a clustering consistency constraint within the cluster space to uncover the complementary information concealed in clustering labels.The proposed X-MVC demonstrates significant performance improvements over multiple competitive approaches in experiments conducted on three challenging multi-view datasets.