Research On Clustering Algorithm Based On Graph Structure

Graph structure can intuitively reflect the correlation between objects.In recent years,clustering algorithms based on graph structure have been widely studied.These proposed traditional algorithms can make good use of the graph structure information to complete the clustering.The clustering results show excellent performance,but these algorithms are highly dependent on the graph structure.With the development of deep learning,graph deep neural network has been proposed.By capturing the graph structure relationship between the objects,the fusion of the neighborhood features is used as the feature representation,which makes features more conducive to clustering.However,the phenomenon of over-smooth feature will affect the clustering effect.Based on the above problems,we make a study on the clustering algorithm based on graph structure.The innovative achievements are summarized as follows:(1)To solve the problem of information loss,an ensemble clustering algorithm based on reconstructed adjacency graph is proposed.In order to reduce the information loss in the mapping process from objects to the categories,the concept of reference set is proposed.The weighted sum of the initial mapping values of the objects in the reference set is used as the modified mapping value from the central object to the category.Thus,the modified mapping matrix obtained effectively increases the amount of information.In order to reduce the sensitivity of similarity measurement to data distribution,Spearman coefficient is used to calculate the similarity matrix between categories.Then the adjacency graph matrix among objects is reconstructed by combining the modified mapping matrix and the category similarity matrix.Finally,hierarchical clustering is performed on the reconstructed adjacency graph.Experiments show the superiority of the proposed algorithm.(2)In order to solve the over-smoothing problem of graph neural network and to ensure that the learned features of the network can satisfy the original near-neighbor relationship,a graph attention clustering network based on local feature constraints is proposed.The features of each layer of the auto-encoder network are implicitly transferred to the corresponding layer of the graph network by using attention network.In this way,the features of graph network can not only integrate the neighborhood information,but also reflect the uniqueness of different objects.Thus,the problem of over-smoothing is alleviated.In addition,the local feature loss function is proposed.By constraining the local similarity among the objects,the local neighbor structure of the dataset is well preserved.Finally,the end-to-end clustering is completed by combining clustering loss and local feature loss optimization.Experiments show that the proposed model can achieve better clustering results.(3)The graph neural network is a cascade structure of multi-layer graph convolution and the graph convolution is a low-pass filter.Therefore,the deep features of the graph network contain less high-frequency information.Based on this,a graph attention clustering algorithm based on feature reuse is proposed.Specifically,the loss of the final feature information is reduced by reusing the deep and shallow features of the graph network.The feature of each layer of the graph network is mapped to the same dimension through the fully connected network.And the feature weight function is introduced to weighted reuse the multi-layer features of the same dimension.In this way,the final feature contains richer information.Finally,the end-to-end clustering is realized by clustering loss.A large number of experiments verify the effectiveness of the proposed clustering model.