Research On Deep Clustering Algorithm Based On Graph Convolutional Neural Network

Clustering is a basic but challenging task in data mining.In recent years,deep clustering method uses deep neural network to extract features from the original input data,and then performs clustering in the embedded feature space.Satisfactory results can be obtained in different data sets,and it has become a hot method of clustering.In practical applications,graph data containing structural information exists widely.The existing deep clustering methods generally extract useful representations from the original features of the data,and pay little attention to the structural information,which limits its application and performance improvement in graph data.Graph convolutional neural network has achieved great success in processing complex graph data by using its modeling ability of structural information and has been successfully applied in various scenarios.However,the performance of graph convolutional neural network depends heavily on the given graph structure,and unreliable graph structure will significantly limit its representation learning ability.In order to solve the above problems,this paper conducts in-depth research on the deep clustering algorithm from the two aspects of fusion of joint coding and optimization of graph structure.The specific work includes:(1)An attribute graph clustering algorithm based on graph convolutional network is proposed to fuse local and global structures.The existing joint encoding attribute graph clustering method uses the simple weighting of the structure encoding feature and the attribute encoding feature to perform clustering.In this method,the local structure is considered through graph convolution,but the global structure is not considered.In order to consider both local and global structures,we propose a joint encoding attribute graph clustering algorithm that fuses local and global structures.Based on the weighted sum,the proposed method uses the first order smoothing features of the weighted sum to construct a kernel matrix to express the global structural features,and then fuses the global structural features and the first order smoothing features to cluster.A large number of experiments on three graph datasets show that the proposed method is always better than the existing typical clustering methods.(2)A graph structure optimization clustering algorithm based on graph autoencoder is proposed.Most graph convolutional neural networks assume that the node relationships described by the observed graphs are complete and accurate,but the graphs from the real world of complex systems are prone to error and may be incompatible with the properties of graph convolutional neural networks.Graph convolutional neural networks that rely too much on the original graph may lead to poor clustering results.We propose a clustering algorithm for graph structure optimization.According to the reconstruction loss of graph autoencoder and the regularization term about graph structure,this method iteratively optimizes the graph autoencoder and graph structure,and realizes the adaptive optimization of graph structure and graph autoencoder for clustering task.This method construct graph structure through feature matrix,which can be extended to clustering of non-graph data.Different from most existing models based on autoencoder,the proposed algorithm does not need pre-training.A large number of experiments show that the performance of the algorithm is good and stable,and is always better than the existing representative clustering methods.