Link Prediction Mechanism For Opportunistic Network Based On Deep Belief Network

The target of link prediction is to estimate the possibility of the existence of links between nodes.Recent years,research in OppNets(Opportunistic Networks)is attracting more and more attention in the academic.Accurate link prediction algorithm can support the research on basic network evolution and upper layer protocol like routes,which can be applied to the recommendation system,public opinion monitoring and other fields.Because of the characteristics of node mobility,intermittent connectivity and limited resources,link prediction in Opp Nets is a hot and difficult problem.Studies in this thesis will expand on link prediction in Opp Nets.According to the characteristics that Opp Nets change frequently over time,a novel similarity index O_AA based on network structure is proposed,which takes the history information and the second order neighbors into account.Then,a link prediction model which is set up by the Conditional Deep Belief Network(CDBN)is applied to extract the variation features of links;The sample space is constructed by the similarity index proposed above through time series method;The network parameters are determined by experiments and the layer-adaptive learning rate is put forward to optimize the training process;The Logistic Regression classifier is adopted to predict links in Opp Nets.MIT Reality and Infocom05 data sets are selected in this thesis.Evaluating indicator AUC and Precision were used for evaluating the feasibility of the proposed O_AA similarity index;groups of contrast experiment are designed for determining the number of layers,sample dimension as well as verifying of models.The experiment results show that the proposed O_AA similarity index can better reflect the change of links in OppNets;through tuning learning rate layer-adaptively,reconstruction error of CDBN goes stable rapidly so that the convergence time is shortened;compared with DBN,CDBN has better modeling ability and the ability to fit the input data,and can get better prediction results.