Link Prediction Approach For Opportunistic Network Based On Recurrent Neural Network

Opportunistic Network(ON)is a new type of mobile Ad Hoc networks that establishes communication through the movement of network nodes without the need of complete communication path between the source and the target node.Due to the characteristics of time-varying topology and node moving frequently,it is more suitable for actual networking requirements in certain application scenarios,and the study of opportunistic network faces many challenges.Link prediction becomes an important topic of the opportunistic network field.The target of link prediction is to estimate the possibility of future links among nodes through known network structures and nodes information.A good link prediction algorithm not only mines the potential relationships of the nodes in the network but also helps to understand the evolution of the network structure,hence,the routing algorithm is supported.According to the time-varying characteristics of the ON,this thesis considers the long-term dependence between the node pairs' historical information in time dimension and its connection state,we propose a novel link prediction approach which is based on recurrent neural network link prediction(RNN-LP)framework.With the help of time series analysis method,we slice the data with fixed period to obtain a series of network snapshots,and analyze the sequence of node pair connection times,which is generated from the network snapshots sequence.The 0-1 test method is adopted to find the chaotic characteristics of time frame connection times sequence,and the chaos theory is applied to determine the length of time frame.We analyze and deal with the relationship between the contact time of the node pair and the start and end time of the time frame,the timestamps of connection and disconnection of node pair are converted into periodical time.The historical information in a unit time frame is defined,and the vector is defined,which is made up of the transformed data and the historical connection information of the pairs in the time dimension,in which a vector sequence of the node pair connection attribute is constructed.The link prediction window is also defined as well,which is the length of vector sequence.Benefiting from RNN in sequence modeling,the RNN-LP model is to learn the correlation between connection state and the vector sequence of connection attribute,and is to extract the inherent law of node pairs in temporal domain,thus to realize the future link prediction better.In terms of sparse degree,some data sets are chosen,such as iMote traces Cambridge(ITC),and MIT reality(MIT).The Area under the Receiver Operating Characteristic Curve(AUC),Accuracy and Precision are adopted as evaluation indices.The Accuracy and Precision of RNN-LP model are computed with different iterations in training,the length of time frame and the length of the input sequence.Compared with similarity prediction techniques of common neighbor(CN),AdamicAdar(AA),Resource allocation(RA),Local path(LP),Katz,the RNN-LP model has a better AUC.Under the different number of test samples,RNN-LP model achieves better stability on ITC and MIT data sets.Experimental results are obtained to reveal that RNN-LP model gives better accuracy than link prediction models of the Linear Discriminant Analysis(LDA),the Support Vector Classifier(SVC),Restricted Boltzmann machine(RBM).