Link Prediction Approach For Pocket Switched Network Based On Firefly Algorithm

The PSN(Pocket Switched Network)is a type of delay tolerant network that utilizes human mobility and opportunistic encounters to exchange message.The topology may change frequently.Therefore the node similarity index based on local information and network topology information is difficult to obtain expected result.In addition,the node movement and data exchanging in PSN are not completely random,but have strongly sociality and regularity because the PSN is composed of people who carry short-distance communication nodes.In this thesis,the regularity of nodes' connection in PSN was analyzed.The community attributes and regularity of the nodes' movement are considered.The node similarity was separated into community similarity and movement similarity.On the one side,constructing the attribute vector of node pair and dividing nodes into different communities by learning automaton.On the other side,calculating the similarity of nodes according to the number of common nodes.Considering the node movement regularity in PSN,a link prediction approach for PSN based on firefly algorithm(FALP)was proposed.Firstly,the node pair feature vector was constructed by the improved similarity index.Benefiting from the autoregressive integrated moving average(ARIMA)model in time series analysis,the ARIMA model is to obtain the node pair feature vector in next moment through analyzing the feature vector of historical moment.Extracting the evolution of the node pair feature vector.Then,a two-classifier is constructed based on the firefly algorithm,which can achive better global optimization.The possibility of a link among nodes in the future can be inferred based on researching the relationship between the nodes' connection states and the nodes' feature vector.AUC,Accuracy,and Precision are adopted as the evaluation indicators.Some datasets are chosen,such as INFOCOM2006 and MIT reality(MIT).Compared with six classical similarity indicators,the improved similarity index have higher AUC and Precision values.The Accuracy and Precision of the FA-LP model are calculated with different initial population numbers,the time frame length,the number of iterations and input sequence length.The experimental results reveal that the FA-LP model obtains better accuracy than link prediction models of the Support Vector Classifier(SVC),Deep Belief Network(DBN),Weak Estimator(WEAK),and Restricted Boltzmann machine(RBM).