Link Prediction For Opportunistic Network Base On GraphSAGE

Opportunistic network is a self-organization network that enables communication through node movement,and it can be used in scenarios such as wildlife tracking,underwater rescue expeditions and network coverage in remote areas.Unlike common networks,the evolution of opportunistic network is extremely complex,featuring sparse node connections and frequent network topology changes.Link prediction is one of the hot issues in the research of opportunistic network.By analyzing the topology and node properties of the network,the features of network evolution are extracted to estimate the possibility of the existence of links between nodes.It provides support for the design of upper layer routing protocols.This thesis presents an opportunistic network link prediction model based on Graph SAGE.Graph2vec model was employed to represent the network,so that its sample entropy and training duration of the model are obtained.Appropriate slicing slot was determined by the entropy and training duration,so as to transform dynamic network data into static-weighted quantified network snapshot information.A representation learning model based on Graph SAGE was constructed to extract the potential features of nodes through multi-order neighbor sampling and feature aggregation.The potential feature similarity of node pairs was calculated,considering the weight of connecting duration between nodes.The connection relationship between nodes and multi-order neighbors was analyzed.The topology similarity of node pairs was calculated according to the local topology information of nodes and the reachable paths between nodes.The potential feature similarity and topological structure similarity were fused by ₂L norm.Historical information was weighted exponentially by improving the trend moving average method,so as to extract the temporally sequence features of fused node similarity.The possibility of connection between nodes in the future was achieved.Experiments were conducted on four real opportunistic network datasets,ITC,MIT,Infocom05 and Infocom06.AUC（Area Under the Receiver Operating Characteristic Curve）,F1-score were employed as the evaluation index.The slice slot was determined based on experiments.The prediction performance of the proposed method was analyzed by setting different time sliding window sizes,node embedding dimension,and time decay factor.And the optimal hyperparameters were determined.The accuracy of the proposed method was achieved by 10-fold cross-validation.The experimental results show that the proposed method has better generalization.Compared with LSTM（Long Short-TermMemory）,SDNE（StructuralDeepNetworkEmbedding）,Struc2Vec（Structural identity to Vector）,E-LSTM-D（Encoder-LSTM-D）,the proposed method has better prediction performance.