Research On Community Detection Method Based On Graph Neural Network

Graph data contains rich relationships and has a wide range of applications in real life.The subgraph structures in which the internal nodes are tightly connected and the external nodes are sparsely connected are called communities.Community detection aims to find the community structure in the graph.The results of community detection can be applied to tasks such as classification,link prediction,recommendation system,public safety,etc.Community detection has important practical significance.Graph neural network is an extended application of convolutional neural network on graph data.It shows superior performance in processing graph data and is widely used in community detection tasks.Existing community detection models based on graph neural network mainly face two problems: on the one hand,traditional community detection models are mostly based on supervised or semi-supervised learning,relying on the label information of nodes in the graph,and the generated node embeddings are more general,failing to contain information for community detection tasks.On the other hand,in the graph,different neighbors of a node have different feature information and importance to the node,and traditional community detection models fail to fully exploit the information between the node and its neighbors,and typically separate community detection from node representation learning.To solve the above problems,this thesis provides corresponding solutions,and the main contributions are as follows:(1)Aiming at the problem that traditional community detection models rely on node label information and the generated embeddings cannot effectively target community detection task,an unsupervised graph neural network community detection model with self-expressive attention(USCom)is proposed,which is a novel model combining graph neural network with unsupervised Learning and introducing the self-expressiveness principle.The model uses both graph structure and node information.First,the adjacency matrix and feature matrix are input into the graph attention network to generate graph embeddings.Then the self-expression principle is used to constrain the generated embeddings to make them more suitable for community detection task.Finally,the model uses a four-layer perceptron to classify node embeddings to generate the community structure of the graph.Experimental results show that USCom outperforms the compared baseline models on four datasets in community detection task.(2)Aiming at the problem that traditional community detection models fail to make full use of the information between nodes and their neighbors,and usually learn community detection and node representation separately,a variational graph embedding community detection model(VGECD)is proposed.The model introduces the graph attention network to perform aggregation operations on neighbor nodes,adaptively assigns different weights to neighbor nodes,and jointly learns the embeddings of node representation and community detection.VGECD first uses an inference model to generate node embeddings and community assignments,and then uses a generative model to reconstruct the graph combining node embeddings and community assignments.Experiments on real-world datasets show that the method has superior performance in community detection task and improves the accuracy of community detection.