Research On Temporal Community Detection Method Based On Heterogeneous Graph Convolutional Neural Network

In recent years,graph neural networks have demonstrated superior performance in representation learning of graph data,which can significantly improve the model performance of downstream tasks such as node classification,link prediction,and community detection.The community detection method discovers the community graph node belongs by analyzing the topology of complex network.These results can be used to predict group behavior,and thus it has become one of the most important research topics.However,the real-world graph data is generally heterogeneous which dynamically varies over time.This greatly challenges on existing graph neural network models.Moreover,graph convolutional networks exist a long-term outstanding issue,i.e.,over-smoothing problem,which needs to be solved.To address these issues,this thesis proposes the temporal-heterogeneous graph convolutional networks(THGCN)and a diverse node sampling strategy based on the determinant point process(Graph-DPP),which are illustrated as follows.Aiming at the task of heterogeneous temporal community detection,this thesis first proposes the THGCN model,which improves the accuracy of community detection by learning heterogeneous temporal graph node representation.The model contains three main components:(1)Heterogeneous graph convolution component,which is responsible for extracting heterogeneous graph node features and network topology information at each moment;(2)Residual compression aggregation component,which is responsible for mining evoluting patterns of temporal graph nodes;(3)Community detection loss function,used to train the feature embeddings of graph nodes.Aiming at the problem of node over-smoothing,this thesis proposes a diverse node sampling strategy based on the determinant point process.This method first designs a feature similarity function and a structural similarity function between any two nodes.Then,the relevance between sampled neighboring nodes and the target node is calculated to build the positive semi-denite similarity matrix.By maximizing posterior probability of the determinant point process and minimizing the similarity between any two sampled nodes simultaneously,we can effectively alleviate the node over-smoothing problem.Finally,the thesis conducts extensive experiments to verify the proposed methods.Experimental results demonstrated that the accuracy of the THGCN model can reach98.93% on DBLP dataset,and the community detection evaluation metric NMI is increased by 16%.In terms of node over-smoothing problem evaluation,on the IMDB dataset,the NMI result of the improved THGCN model based on Graph-DPP is increased by 10%.To sum up,the model proposed in the paper can achieve better community detection results on heterogeneous temporal graph data,and the diversified sampling strategy proposed in the thesis can further improve the model performance.