Multi-view Clustering Based On The Similarity Graph Fusion

With the generation of massive multi-view data,how to analyze and integrate these heterogeneous multi-view data becomes an urgent need.Multi-view clustering aims to use the common and complementary information of multi-view data to cluster.Multi-view graph clustering is one of the most representative methods of multi-view clustering.The multi view graph clustering firstly constructs the similarity graph for each view,then merges multiple similarity graphs as a unified graph,and then completes the clustering on the unified graph.We are deeply understanding the works of multi-view graph clustering,and the key to these algorithms is to measure the similarity accurately.However,existing multi-view graph clustering methods are not accurate enough or can not be used to different data types well when constructing similarity graph.In addition,most of these algorithms do not consider the importance of different views in the process of graph fusion,thus affecting the stability and accuracy of clustering.For these problems,this paper studies the following contents:(1)The research of multi-view graph clustering algorithm.Firstly,we introduce the theoretical knowledge,Laplace matrix,high-order similarity and spectral clustering algorithm.Then,we introduce some representative multi-view graph clustering algorithms and analyze the advantages and disadvantages of the existing methods.(2)In view of the single type of measure used in the multi-view graph clustering method,which leads to the lack of accuracy and generalization ability of similarity measurement,we propose a similarity measurement based on conditional probability and improve the accuracy and generalization ability of measuring similarity by using multi measure.We design an effective and adaptive multi metric similarity fusion model.The model automatically learns the weight of each view and each measure,and the graph fusion and clustering promote each other,and directly learn the clustering results.The comparison and ablation experiments prove the effectiveness and superiority of our model.(3)For most of the multi-view graph clustering methods,we propose a new high-order similarity,which is based on the first-order similarity and limited accuracy.In addition,in order to maximize the global information of each graph,we integrate the clustering information of each graph into the corresponding high-order similarity matrix.We propose a similarity graph fusion model,which can automatically learn the weight of each similarity graph,and extract the common information and complementary information of all views.We use this model to perform subtyping experiments in cancer multi-omic data set to verify the clustering performance of the model.