Research On Dynamic Link Prediction Based On Improved Graph Convolution

Link prediction is one of the important means of network analysis,which can be widely used in many practical scenarios,such as online social networks,disease transmission,etc.Its primary task is to predict the possible links between entities in the network.However,networks in the real world tend to evolve over time.Dynamic link prediction has been paid more attention in recent years because it is more effective to capture the evolution law of the network.However,most of the current researches fail to achieve satisfactory prediction performance,and the model itself is not scalable for complex network data,which will lead to heavy training tasks and long training time.Aiming at the computational memory and communication problems faced by large-scale link prediction model,this paper improved the mainstream Distributed solutions,and proposed a Distributed parallel training model of GCN(DPGCN)based on random boundary node sampling.In view of the excellent performance of gated cycle unit(GRUs)in processing temporal network data,DPGCN model is proposed by combining DPGCN model with GRUs,and is applied in the field of dynamic link prediction.The thesis focuses on the following three aspects: Firstly,by analyzing the performance bottleneck of distributed solution,a random boundary node sampling algorithm is designed,which can divide complex network into several subgraphs,and use node sampling to iteratively reduce the number of boundary nodes in the subgraphs,and then conduct parallel graph convolution training for the subgraphs,so as to save the model communication overhead.Secondly,in the parallel training stage,the sparse matrix and the calculated data locality are used to extend the distributed training process,and the training performance of the model is further optimized.Finally,a stacked gated cycle unit is used as the encoder to extract the time characteristics of the input sequence,and the attention mechanism is introduced to weight the historical time series and dynamically adjust the importance of network information contained in different time slices,so as to improve the overall prediction effect of the model.In order to evaluate the performance of the model in practical application,based on the two models proposed in this paper,two link prediction scenarios of static network and sequential link network were set respectively,and a total of five groups of comparison experiments were designed to test the prediction effect on large-scale real data sets.The experimental results show that in static link prediction scenario,the DPGCN model reduces the training time by 27% on average compared with other link prediction models.In dynamic link prediction scenario,compared with other sequential link prediction models,DPGCN-GRU model adopted in this paper has high accuracy,strong stability and other performance,and has broad application prospects.