Research On Methods For Link Prediction In Social Networks

With the rapid development of internet techniques, SNS(Social Network Service)becomes a part of many people’s daily life. As a novel kind of internet service, SNS provides many functions which are similar in real social environment, and such functions could attract new members to register for SNS with different interests. Most of these SNSs provide many kinds of functions for member interactions. For example, one function may allow members to mark others as friends or foes, another function may let a member become certain administer by election, and another function may allow members to discuss or comment other members’ publishing or opinions. The rich data of these members’ interactions is accumulated by real social member themselves, so it could reflect some common scene in the development of human society. As a result, if we could analyze the data of member interactions properly, it would be very important for the researches of social networks, internet development, and other related business.Most of the interaction among members could be represented as ’link’. If the social member is represented as ’node’ in the graph, the interactions among them are represented as different kinds of links, each of which links two nodes, and such graph is named as’ link network’. By analyzing the structure of link network, the link prediction problem aims to estimate the possible values of the links. In this condition, with the research of link prediction, it would be great help for the research of member interaction. However,most common link prediction method would be unavailable in the condition that the link network structure is unobserved, or make prediction for new joining members. And the links in link network is not static or discrete. The dynamic emerging of links is affected by many factors, but most researches did not take them into account. If kinds of information sources could be used to solve link prediction, the performance of link prediction methods could be improved. In such condition, it could help the analysis of social member interaction, and could provide better services for members.This dissertation focuses on methods for link prediction in social networks. The research extends with finding available features and suitable methods for solving link prediction. The researches in this dissertation includ link network structure based link prediction and feature analysis; improving the link prediction and features’ performance by deep learning method. In this dissertation, user generated context and sentiment lexicon are used to improve prediction performance, and the attempt of modelling the static link network as dynamic sequence behaviour is performed. With more features are taking into account, the multimodal approaches are used to make full use of them. In detail, the main research content contains the following four parts:(1) The link network, as the basic research object in link prediction, is worthy for detail study and analysis. By analysing the structure of link network, we extract as many features as possible. Because the scale of link networks is always very large, it would cost much time to finish extracting features. The optimization method for feature extracting is proposed, and it has high efficiency. Based on linear support vector machine, the link network structure based link prediction method is proposed. By analysing the model and connecting with certain social theory, how each feature effects the link value is discussed.(2) This dissertation focus on solving certain problems in link prediction, some available solutions are proposed. In order to solve the cold start problem, by using some sentiment analysis method, the user generated context and sentiment lexicon are taking into account to solve link prediction. Based on features from multi-sources, the improved methods for user similarity metrics based prediction and machine learning classifier based link prediction are proposed. In order to solve the problem of only considering the link as static and discrete prediction unit, user link behaviours are changed to dynamic and continuous sequence behaviour. Based on modelling the sequence behaviour, the link prediction method for sequence behaviour is proposed, and the prediction method for the final result of sequence behaviour is also proposed.(3) There should be rich information of member behaviours hidden in the link network structure. In order to model such information, the deep learning approaches are used. By analysing the ability of RBM(Restricted Boltzmann Machine), three different functional DBN(Deep Belief Network) are proposed. Based on these DBN structures, the deep learning approaches based unsupervised link prediction method, feature represent method, and DBN based link prediction method are proposed. In order to save training time, an optimized learning strategy is proposed. And the generalization across datasets from social media with different interests is performed. The experiment results show that the deep learning method could make better use of link network structure features, and the prediction performance is improved.(4) In the condition that different sources of data are available in some social networks, the different modal features should be well used to solve link prediction problem.In order to learn the joint distribution of features from different modals, two kinds of MDBN(Multimodal Deep Belief Network) structure are designed, and the multimodal learning approaches based link prediction methods are proposed. With the research of DBN, focusing on label classification task and sample reconstruction task, this dissertation proposes the learning methods for two different functional DBN. By building the MDBN with different functional DBN, the performance of link prediction is improved.At the same time, the method for generating the missing part of features is also proposed.