Research On Self Supervised Community Detection Algorithm Based On Graph Neural Network

As a typical representative of non-Euclidean structure data,graph data can well reflect the abstraction of the real world.In recent years,the research on graph data has become a hot field and many tasks have been widely used in real scenes,such as link prediction,multi-hop reasoning,and community detection.Community detection is one of the typical applications of graph data analysis.It aims to mine the hidden relationships among nodes from complex graph structure networks,such as finding people with similar hobbies from social networks or dividing weapons with similar attributes in the military knowledge graph.However,the representation of graph data has become more and more complex and the data scale has become larger and larger,which leads to poor performance and deficiency with traditional community detection methods.Inspired by the successful application of deep learning in computer vision and natural language processing,scholars are committed to combing deep learning with graph representation learning,and then proposing Graph Neural Networks(GNNs).GNNs have achieved superior performance in graph data analysis tasks,such as community detection.However,there are still some limitations: 1)due to the complexity of constructing graph data,some graph data lack enough labeled data,which makes GNNs can only utilize few labeled data but ignore most of the unlabeled data.2)most of GNNs only learn graph data from a single view,ignoring the local and global information of graph data.Especially for the community detection task,the nodes in the community and the overall distribution of the community play important roles.3)some GNNs absence a reliable explanation for the community detection task-oriented objective function in theory.Especially in the unsupervised scene,the lack of labeled data makes the objective function of GNNs irrelevant to community detection.To address the above issues,this thesis designs the new graph neural networks for community detection in semi-supervised and unsupervised scenes,respectively.Specifically,in the semi-supervised scene,a new end-to-end Iterative Feature Clustering Graph Convolutional Network(IFC-GCN)is proposed.IFC-GCN constructs the Iterative Feature Clustering(IFC)module to enhance the representation learning ability of the classical Graph Convolutional Network(GCN).In the IFC module,a discrimination matrix is developed based on pseudo labels and clustering labels.The pseudo labels are obtained by GCN pre-training and the clustering labels are obtained by performing clustering on the hidden layer features.The similar relationship between node pairs can be identified through the discrimination matrix,and then an optimization model is constructed to iteratively optimize the node representations.In addition,an EM-like framework is designed for model training,which improves the network performance by rectifying the pseudo labels and node representations alternately.In the unsupervised scene,a Contrastive Learning Framework with Multi-Granularity Feature Interaction(CL-MGI)is proposed.In CL-MGI,the contrastive learning loss function Info NCE is introduced,and its applicability to community detection is analyzed theoretically.Then,two independent contrastive modules are constructed from node-level and graph-level respectively.The node-level module learns fine-grained node feature information,and the graph-level module learns coarse-grained community distribution information.Further,an adaptive feature fusion method integrating graph topology information and node features information is proposed to select contrastive sample pairs in the node-level module.This method can select unbiased positive and negative sample pairs to prevent the local overfitting of node features.More importantly,a temporal entropy-based metric is introduced to evaluate the sample quality and realizes the multi-granularity feature interaction in a co-teaching manner,thus realizing the fusion of multi-granularity feature information.In this thesis,the graph neural networks designed for semi-supervised and unsupervised scenes are essentially self-supervised learning methods.They make fully utilize the information of the data itself to learn node representations.Full experiments under the public data sets of semi-supervised and unsupervised scenes are carried out.The experimental results demonstrate that the proposed methods show superior performance in both the accuracy of the community detection task and the generalization of the model.