Research On Temporal Network Link Prediction Method Based On Deep Learning

With the development of large-scale networks,temporal link prediction has become one of the hot research topics in the field of complex networks,especially in the field of economics,sociology and biology.Compared with link prediction in static networks,link prediction in temporal networks is more challenging because of the complex and dynamic nonlinear spatial-temporal correlation.At present,most of research related to link prediction focuses on static networks,but their prediction performance are not as expected in temporal networks.Therefore,the research on link prediction of temporal networks has become an important task in the field of network representation learning.The difficulty of the research is how to capture the spatialtemporal features of the networks and fuse them effectively,and then predict links in the networks.Therefore,this paper conducted in-depth research on the link prediction problem based on temporal networks.The main contents are as follows:In this thesis,we propose a temporal link prediction model based on spatialtemporal interaction fusion(STIIF).The model consists of three modules: SpatialTemporal Features Encoding Module,Spatial-Temporal Collaborative Interaction Module,and Spatial-Temporal Feature Fusion Module,which can fully exploit the spatial structure and dynamic characteristics of temporal networks to improve the accuracy of link prediction.Firstly,the Spatial-Temporal Features Encoding Module learns the spatial features of each network snapshot through Chebyshev Graph Convolution.Then the Spatial-Temporal Collaborative Interaction Module interactively enhances temporal and spatial features through a collaborative interactive attention mechanism.Finally,Spatial-Temporal Feature Fusion Module uses gated recurrent unit to capture hidden temporal features and evolution patterns in temporal networks,and fuses them with spatial features,so that predict the structure of the network snapshot next moment.To verify the model’s effectiveness,simulation experiments are carried out on four datasets of UCSB,Twitter Tennis,College Msg and England Covid19.The experiment results show that the general predictive ability of the STIIF model on AUC,KL divergence and Mismatch Rate is better than other similar algorithms.Although the STIIF model has a good performance,it is ineffective on large-scale sparse networks.Therefore,we propose a temporal link prediction model based on variational diffusion graph convolution(VDGC).The model aims to extract and aggregate spatial-temporal features in large-scale sparse networks to obtain more accurate representation of nodes in the network.The VDGC model amplifies the relevant features of aggregated nodes in the network through Diffusion Graph Convolution,and suppresses them of scattered nodes.It improves performance of capturing spatial features in sparse networks.The model learns the probability distribution of node features and sample the final node representation through the Variational Graph Auto-Encoder,which enhance the robustness of node representation and improve the predictive performance of the model.The VDGC model is simulated on three real datasets of Twitter Tennis,College Msg and Fb-Forum.Experimental results demonstrate the superiority of the model.And it is more suitable for large-scale sparse networks than the STIIF model and other baseline methods.