Research On Multi-view Clustering Algorithm Oriented To Attribute Graph

With the rapid development of sensor technology and big data technology,the data volume of various domain is growing exponentially.These data can be collected,processed and stored from various modalities,presenting characteristics like multi-view,graphstructured and unlabeled.Multi-view clustering algorithm aims to learn the potential relationship between views by making full use of the property of different views in an unsupervised manner,so as to obtain more complete information and effectively divide the data into disjoint clusters,and further handle the problems existing in many real-world scenarios such as data mining and analysis.Multi-view clustering algorithms have developed rapidly in the past decades.Although these methods have shown excellent performance in real-world tasks,there are still some problems and challenges.Firstly,traditional multi-view clustering methods can only deal with feature data or graph structure data,but cannot capture complete data information by using both the feature information and topological structure at the same time.Secondly,the heterogeneous information in different views and the noise in the original data hinder us from accurately capturing the clustering structure of all views,resulting in the degradation of clustering performance.Finally,existing methods only utilize simple clustering loss constraints or combine embedding representations when computing consensus representations of different views,and do not fully explore the potential relationship and consistency of view embeddings,which limits their ability to handle complex situation involving multiple graphs and attributes.To address these problems,this thesis studies the generic and robust multi-view clustering algorithm oriented to attributed graph.The specific research is as follows:(1)This thesis presents a Smoothed Multi-View Subspace Clustering algorithm.This method constructs a graph for each view,and introduces graph filtering technology to obtain a smooth representation of each view.Furthermore,based on multi-view subspace clustering,complementary information and clustering structure between views are mined.The graph filtering not only preserves the topological structure information well,but also facilitates a clustering-friendly representation.The experimental simulation shows that the algorithm achieves good results,and verifies that the feature representation after graph filtering has high separability,which is beneficial for downstream clustering tasks.(2)The algorithm proposed in(1)is only capable of handling multi-view attribute data.To effectively process multiple graph data with attributes,this thesis proposes a Scalable Multi-view Clustering with Graph Filtering and Anchor Graph algorithm.The algorithm enhanced(1)by introducing a scalable graph filtering that selects an optimal filter for different datasets.In addition,a node importance sampling strategy is adopted,which can capture the topology structure of multi-view attribute graph data efficiently and reduce the time complexity of multi-view sbuspace clustering.Experimental results demonstrate that the proposed algorithm significantly improves the clustering performance and shortens the running time.(3)Existing methods do not fully explore the potential relationship and consistency of view embeddings,so that they cannot solve complex data with multiple graphs and attributes.Therefore,this thesis introduces a generic and effective clustering algorithm framework,Multilayer Attributed Graph Contrastive Clustering Network.This method employs a multi-layer graph attention autoencoder along with contrastive learning mechanism to capture consistency information across different layers,and then a self-supervised clustering strategy is utilized to iteratively enhance node embedding and clustering distribution,and finally obtains clustering results with high confidence.Extensive experiments on benchmark datasets demonstrate the superiority of the proposed algorithm,which has powerful generality and robustness.