A Research On Deep Multi-View Clustering Model Based On Graph Representation Learning

Clustering is an important topic in artificial intelligence that has long been a subject of extensive research.In recent years,due to the richness of multimedia data in the real world,multi-view clustering has become one of the hot research areas.The study of deep multiview clustering effectively addresses some issues in traditional shallow multi-view clustering,such as high complexity,low representational capacity,and difficulties in feature extraction.Deep multi-view clustering leverages the powerful representational and feature extraction capabilities of deep learning,driving new developments in multi-view clustering research.However,existing deep multi-view clustering methods struggle to handle multi-view data clustering in non-Euclidean spaces,specifically clustering on graphstructured data from multiple views.On the other hand,graph structures are common data structures in the real world and are widely used in domains such as chemical molecular graphs,biological protein graphs,social network graphs,and citation graphs.In recent years,several graph representation learning methods have emerged with the aim of uncovering the topological structures within graph-structured data.These methods effectively learn the topological structures in graph data through techniques such as semi-supervised learning and graph neural networks.However,when faced with graph-structured data that simultaneously contains multiple views,these methods perform poorly.For the same dataset,multiple different graphs can be constructed.For example,in a social network dataset,one graph can be constructed based on user relationships,while another graph can be constructed based on shared interests.Further research is needed on how to leverage the topological consensus and complementarity among multiple graphs and how to eliminate noise using multiple graphs.Therefore,this thesis focuses on the research of deep multi-view clustering algorithms based on graph representation learning.Starting from graph representation learning,this thesis first customizes graph neural networks by extracting high-level semantic information specific to different downstream tasks,proposing a path reweighting customized graph neural network.This addresses three inherent problems of traditional graph neural networks: 1)Traditional graph neural networks often suffer from overfitting to the dataset due to the lack of label information in the graph-structured data during the training process; 2)Graph neural networks are prone to introducing additional noise during the message passing process,resulting in poor robustness against data attacks; 3)The inherent oversmoothing problem of graph neural networks arises from the multi-step message passing process,leading to overly smooth node embeddings that are difficult to distinguish.Subsequently,based on the path reweighting mechanism,the thesis attempts to integrate the global self-attention mechanism with the hierarchical attention mechanism of graph neural networks,proposing a shared attribute multi-graph clustering algorithm based on the global self-attention mechanism.This combines graph neural networks with multi-graph structures in multi-view clustering,addressing three major challenges in multiview clustering tasks for graph data: 1)The structural information of multiple graphs exhibits significant differences,and this approach resolves how to leverage these diverse pieces of information; 2)Some graph structures contain many noisy edges,and this approach resolves how to mitigate these noisy relationships; 3)Additionally,it explores how to better utilize complementary information among multiple graphs.Finally,this thesis delves deeper into the research on multi-view graph clustering tasks,focusing on two characteristics of multi-view clustering: consensus and specificity.By designing a variational graph generator to directly generate sparse graph structures,this thesis emphasizes the extraction and fusion of consensus and specificity information of topology and features within multi-view graph data,as well as the relevance between the model and clustering tasks.This makes a new contribution to the research of deep multi-view clustering algorithms based on graph representation learning.