Temporal Link Prediction Model Of Weighted Dynamic Network

In the real world,network structures naturally exist in various complex systems in the real world,including social networks,shopping networks,co-authoring networks,Ad-hoc networks,and computational biology.Many problems in these complex systems can be reasoned by network structures.Traditionally,network analysis methods mostly focused on static networks.However,many real-world networks show various dynamic behavior,including topological evolution,feature evolution,and feature diffusion.In fact,dynamics is an important factor that determines the performance of most network systems.Therefore,the research of network dynamics is regarded as a key problem that needs to be explored deeply,and the link prediction task of dynamic network is considered to be an effective means to study this problem.Although the research on temporal link prediction of dynamic networks has increased rapidly in recent years,it is still a challenging task to combine the temporal information with the network topology information and predict the link in dynamic networks because there are a lot of temporal information that is difficult to capture in dynamic networks.In this thesis,we propose a dynamic network link prediction model named GARG,which based on generative adversarial learning.This model combines graph convolution network(GCN),gating recurrent neural network(GRU),attention mechanism and generation antagonism network(GAN),which can make full use of the dynamics,topology and evolution mode of dynamic network to improve the performance of temporal link prediction.Concretely,GARG first utilize GCN to explore the spatial characteristics of each single snapshot and then employ GRU to capture the hidden temporal features and evolution patterns in the dynamic network,combined with the self-attention mechanism to further enhances the understanding of the most relevant time points of each network snapshot.Moreover,GAN is used to enhance the ability of the GARG model to learn dynamic network features and generate high-quality predictive networks,which effectively solves the problem of link sparsity and wide range of link weights in dynamic networks.To verify the model's effectiveness,simulation experiments are carried out on four datasets.Experimental results show that the comprehensive performance of GARG on RMSE,KL divergence and Missing Rate is better than that of the baseline method.Although the GARG model performs well in small and medium-sized networks,it is difficult to expand to large-scale networks due to the high computational complexity and memory usage.Therefore,this paper proposes a dynamic network temporal link prediction model(VTSA)via recurrent variational graph convolutions.This model aims to exploit the rich temporal and spatial information found in large dynamic network to capture the topology and evolution mode of the network,and to learn the representation of nodes in the network.The VTSA model exploits hierarchical recurrence at different depths within the network to enable exploration of changes in temporal and spatial.And share parameters at multiple time steps to achieve the purpose of reducing the overall parameters of the model,so that the VTSA model can be more flexibly extended to a larger network.Through the simulation experiments on three real datasets,the VTSA model shows good performance.The comparison with the GARG model proves that it is more suitable for large-scale dynamic networks.