Link Prediction Methods Based On Evidence Theory

With the development of information technology,the research of complex network analysis has been continuously deepening.As an important research direction of com-plex network analysis,link prediction can explore the potential links in the network and has a broad of applications.In recent years,researchers have proposed many structural similarity-based link prediction methods.These methods have low time complexity and high interpretability,but most of them are limited to considering one network structure and thus ignore the structural differences between networks.Therefore,the performance of these methods is unstable on diverse networks.To address this problem,our thesis proposes two new methods,namely IMP?DS and EKPNN,by fusing different network structures using the DS evidence theory.(1)IMP?DS: the method by fusing the clustering coefficients of nodes and links.Node clustering coefficient and asymmetric link clustering coefficient are commonly used network topological structures,which measure the possibility of the link between two nodes from different perspectives.In this thesis,we fuse these two clustering co-efficients by DS evidence theory,and compute the similarity scores of candidate node pairs by employing the IMP model.Due to the fusion of two clustering coefficients,the IMP?DS method has better prediction accuracy and adaptability than the methods that only use a single network structure.The experimental results on nine real networks show that the IMP?DS method has better prediction performance than the compared methods.(2)EKPNN: the method based on evidences of K-pairs nearest neighbors.To fuse more network structural information to improve the performance of link prediction,we propose the EKPNN method by combining the DS evidence theory with KNN algorithm.First,we use multiple network structures to construct the feature vectors of node pairs.Then,we search the k pairs nearest neighbors of a potential link,and treat each pair of the nearest neighbors as an evidence source.Finally,based on DS evidence theory,we fuse the information obtained from the k evidence sources to compute the similarity score of the potential link.To verify the prediction accuracy and applicability of our method,we conduct experiments on twelve real networks.The experimental results show that(1)the EKPNN method has superior prediction performance,and(2)fusing multiple network structural features by DS evidence theory can effectively improve link prediction effects.