Research Based On Representation Learning Of Temporal Attributed Graph And Clustering Method

Graph structure data refers to the data formed in topological space according to the relationship between things.It widely exists in people’s real life,and is widely used in social network analysis,disease transmission analysis,public opinion survey and other fields.Graph data can be divided into temporal graph and static graph according to topological structure and whether node attribute information evolves over time.At present,the research on graph structure data mainly focuses on static graph,such as graph convolution neural network and graph attention network.However,in the real world,the structure and node properties of graph networks usually change with time,so the research method based on static graph can not meet the needs.In addition,there are three major problems in the existing representation learning and clustering methods of time series attribute graphs:(1)the mainstream research is usually supervised learning based on labeled data sets,but in reality,the labeling cost of data sets is high,so there are great application limitations.(2)This kind of method is prone to the problem of over smoothing of the representation convergence of prediction nodes,which leads to the limitation of data feature extraction.(3)The existing clustering methods of time series attribute graph usually separate graph representation learning from clustering algorithm,so it is unable to combine deep representation learning with clustering,which makes it lack of modeling ability of graph temporal evolution process.In view of the above problems,this thesis mainly completes the following two aspects of research.In this thesis,an unsupervised representation learning model based on graph self encoder is proposed.Firstly,the residual network is used to optimize the graph self encoder to alleviate the over smoothing problem.Secondly,the damaged graph is reconstructed for unsupervised learning of link prediction to remove the dependence on the labeled data.Finally,the time series of graph network is modeled by the gating cycle unit.Experimental results show that the performance of the model is better than that of the traditional model.In this thesis,we propose a deep clustering model combined with temporal evolution.The model uses double self encoders to fuse the deep clustering network,which makes the graph representation learning and clustering learning process simultaneously,improves the representation ability of the model in clustering and uses the temporal evolution information sharpening layer to assist prediction,enhances the modeling ability of the model for the graph evolution process,and improves the performance of graph representation learning results including time series information on clustering tasks.Experimental results show that the effect of the model is significantly improved compared with the traditional clustering method.