Research On Methods Of Link Prediction In Social Networks

As one of the fundamental research problem in network data analysis,link prediction has great significance in exploring the law of network evolution and in completing the missing data;it is also of great significance for many applications,such as recommender systems,e-commerce,scientific collaboration,etc.In the Web2.0 age,the emergence of social media platforms(such as Wechat,Weibo,Facebook,and Twitter)have provided unprecedented opportunity for us to explore in-depth the social behaviour,network evolution,and organization pattern and so on.For link prediction,many important methods have been proposed based on the topological information,the non-topological information or the information fusion.Despite these significant advances,many challenges in link prediction remain to be solved due to the complexity of network structure,the multi-source data,the cold-start problem,and the dynamics in social networks.The thesis aims to deal with the challenges of the link prediction problem in different social data environment and to propose corresponding and effective models and methods for link prediction.The main results obtained are as follows.(1)For modelling the underlying topological semantic in a social network,we propose fusion probability matrix factorization models by fusing the topological metrics between nodes.In the models,we consider not only the symmetric node-to-node topological metric but also the asymmetric node-to-node topological metric and both of them.The extensive experiments show that the proposed models give better link predicting performances than the other related comparison methods.(2)For modelling the rich text content published by social network users,we propose fusing probability matrix factorization models by fusing the topic semantic between users.In this work,we define a kind of user topic similarity(TS)based on KL divergence and a kind of user topic inclusion degree(TID)based on the characteristics of information spreading in social networks.In a probability matrix factorization framework,we build the fusion probability matrix factorization models by fusing the TS and the TID respectively.The experimental results demonstrate that the proposed method is more effective than the other related methods in solving the link prediction problem.(3)For solving the cold-start problem,we propose a cold-start link prediction framework.In the framework,the existing network nodes are first represented in a latent-feature space;then we propose a L-logistic mapping model,which map the cold-start nodes into the latent-feature space;finally the cold-start link prediction problem can be solved in the unified latent-feature space.The experimental results show that L-logistic mapping model performs better in establishing connections between topological and non-topological information,and the proposed approach is very effective in solving cold-start link prediction.(4)For modelling the evolution of social network,we propose a temporal probabilistic matrix factorization model.In the temporal probabilistic matrix factorization model,we present a near-dependence based probabilistic generative strategy,and the snapshots of an evolving network can be modelled naturally in a unified probabilistic generative framework.In the framework,the network at each timestamp can be represented in a latent-feature space,and the latent-feature representations at any two adjacent timestamps are similar.Our experiments show that the link prediction method based on the temporal probabilistic matrix factorization model outperforms the other representative methods.In conclusion,this thesis proposes corresponding and effective models and methods for dealing with the link prediction problems under different social data environment,and these studies provide some important contributions for link prediction in social networks.Meanwhile,these contributions have enrich the system of models and methods in link prediction.Moreover,the research results provide new technology support for some applications,such as recommender systems and scientific collaboration.