Link Prediction Method For Opportunistic Networks Based On Random Walk And Deep Learning

Opportunistic network is a type of mobile ad hoc networks that does not require a complete link between the source and target node,and the communication between nodes is carried out through the encounter opportunities brought by their movement.Messages are transmitted between nodes in the form of"store-carry-forward"until they are transmitted to the target node.Link prediction is one of the hot spots and difficult problem in the research of opportunistic networks.Link prediction is to evaluate whether there is a link between nodes based on the known links and the attributes of nodes.An excellent link prediction method can not only discover the law of link change,but also assist researchers to further understand and anal yze the topology changes of the network,thereby providing theoretical support for designing opportunistic network routes.According to the characteristic that the topology of opportunistic network changes frequently with time,this thesis proposes a novel link prediction method for opportunistic networks based on random walk and deep learning.Based on the second-order neighbors and the interaction of the nodes,we reconstruct the Markov transition probability matrices of network separately,and propose two improved random walk with restart similarity indexes(IRWR):IRWR~n andIRWR~i;We cut the opportunistic network to obtain a series of time-series network snapshots,and the similarity matrixs of the snapshot are constructed based on two IRWRs,then we construct a sample set according to the impact of historical state of the link on its connection state.We utilize the advantages of the deep belidf network(DBN)model in pre-training and feature extraction,and propose a link prediction model based on improved random walk with restart and deep belief network(IRWR-DBN)to learn the relationship between the link changes and its historical information,extract the intrinsic characteristics of the link changes,and then predict whether the node pair will be connected in the future.In the paper,we select three real datasets with different network sizes and connection sparsity,ITC,HYCCUPS,and MIT as experimental datasets,and we select precision and AUC to evaluate the performance of the IRWR-DBN model.Based on the sample data constructed based on theIRWR~n andIRWR~i indexes,we set different parameters such as the length of the slice time and the length of the input data,and calculate the AUC and Precision of the IRWR-DBN model.Then,we can determine the optimal model on each dataset according to the results of comparative experiments;Compared with the link prediction methods of CN,AA,Katz,RA,RWR,CNN,RNN-LP and E-LSTM-D,the proposed IRWR-DBN model has better performance than the classic similarity index model,and The IRWR-DBN model is more stable and accurate than deep neural network models such as CNN,RNN-LP and E-LSTM-D.