Research And Application Of Node Clustering Algorithm Based On Graph Attention Mechanism

In reality,there are various complex networks,such as biomedical networks,information technology networks,and social networks.In recent years,the study of complex networks has become one of the hot topics in interdisciplinary research across various fields.The main task of complex network clustering is to discover the real network clustering structures in complex networks.The importance and application value of complex networks in current research lies in their positive role in predicting network behavior,analyzing network topology,and exploring potential network functionalities.Recently,network embedding techniques have been integrated into node clustering tasks in two ways to capture the complex relationships between nodes.In the two-stage approach,node embedding vectors are generated and clustering algorithms are applied to obtain node clustering results.In the one-stage approach,node embedding vectors and clustering results are obtained simultaneously by optimizing a mixed objective related to the relationship between nodes and cluster centers.However,the disadvantage of the two-stage approach is that the learned embedding representation may not lead to the best node clustering results,and the node clustering results cannot optimize the process of graph embedding learning.The one-stage approach ignores the position of nodes(in the core or on the boundary)in clustering and the impact of node feature information on clustering generation.Existing algorithms usually only consider the topological structure information of nodes and do not fully utilize node feature information,resulting in the waste of node feature information.In reality,nodes with similar structures often have similarities in feature information as well.By combining deep learning and graph embedding learning,this paper proposes a node clustering algorithm based on graph attention mechanism to effectively address the aforementioned issues.The research work and contributions of this paper are as follows:(1)An improved algorithm called DAEGC-X~2(Graph Clustering Using A Deep Attentional Embedding Approach Enhanced by Taking Full Advantage of Nodes Features)is proposed,based on the enhancement of node feature information in the original DAEGC algorithm.The DAEGC algorithm presents an objective-oriented deep learning method that focuses on graphs with node feature information to explore both topological and feature information.By using an attention network to capture the importance of neighboring nodes to the target node,the DAEGC algorithm encodes the topological structure and node features of the graph into compact representations and reconstructs the graph structure through training an inner product decoder.Additionally,the DAEGC algorithm supervises the self-training graph clustering process by generating soft labels for graph embedding itself,fully optimizing the clustering results through iterative iterations.However,the DAEGC algorithm does not fully utilize node feature information in networks with node feature information.The proposed DAEGC-X~2 algorithm strengthens the acquisition of node feature information through different methods.Firstly,the algorithm constructs a node feature similarity matrix to capture the feature similarities between each node and other nodes,integrates the feature similarity into the topological structure as part of the calculation of the graph attention mechanism coefficients,and fully utilizes node feature information to enhance the performance of node clustering.Secondly,the DAEGC-X~2 algorithm constructs a node feature similarity network based on the constructed node feature similarity matrix to capture more node feature information between nodes.In the node feature similarity network,the edge weights between nodes represent high-order neighborhood information,which can comprehensively describe the relevance between nodes.The proposed DAEGCX~2 algorithm is experimentally evaluated on multiple real networks and compared with other graph neural network algorithms to assess its performance using node clustering evaluation metrics.The experimental results demonstrate that the proposed algorithm outperforms other comparative algorithms in node clustering tasks,indicating the superiority of the deep attentional embedding approach enhanced by taking full advantage of nodes features.(2)An enhanced version of the Biased Random Walk Node Clustering Algorithm based on Attention Mechanism(AGRWNC)is proposed.The original BRWCD algorithm is a node clustering algorithm based on biased random walks.It enhances the clustering partition by constructing a topological weighted degree matrix on the topological structure and adding node feature vectors as attribute nodes to the original network to construct an attribute-enhanced network.The probability transition matrices from topological nodes to attribute nodes,from attribute nodes to topological nodes,and from attribute nodes to attribute nodes are constructed to further enhance the topological weighted degree matrix.The node embedding representation is obtained by combining these three matrices using random singular value decomposition,and the final node clustering results are obtained using the K-Means algorithm.However,the BRWCD algorithm does not fully utilize the topological information of nodes in the calculation of the topological weighted degree matrix,does not consider the correlation between node features when using node feature information,and cannot simultaneously optimize the node embedding representation and node clustering due to its one-stage process,resulting in suboptimal results.The AGRWNC algorithm fully captures the topological information of nodes on high-order neighborhoods by using the strategy of graph attention mechanism,assigning different weights to different neighboring nodes.Secondly,to fully utilize node feature information,this paper distinguishes the influence of different attributes on nodes by calculating the probability transition matrices from topological nodes to attribute nodes,from attribute nodes to topological nodes,and from attribute nodes to attribute nodes in the attribute-enhanced network,integrates the four probability transition matrices,and obtains the preliminary node embedding representation using random singular value decomposition.Finally,the AGRWNC algorithm uses a self-optimizing clustering module to optimize node embedding representation and node clustering results in both directions,making the learned node embedding representation better adapt to node clustering tasks.The experimental results demonstrate that the AGRWNC algorithm overall outperforms other comparative algorithms,demonstrating the efficiency of the biased random walk node clustering algorithm based on attention mechanism.