Research On Key Technologies Of Construction And Analysis Of Social Relations

Online social networks are changing people's life,and these changes are propagated through social relations.Social relations are the relationships among all the users in a social network.Studying the construction of social relations helps us understand the generation and the evolution of social networks.Studying the analysis of social relations helps us analyze the social network based on different properties of social ties.This dissertation proposes four research problems from the aspects of both the construction and the analysis of social relations.It studies how to construct social relations among new users and old users as well as among new users themselves.It also studies how to represent social ties and how to analyze the directionality of social ties.The primary contributions are summarized as follows:1.To construct social relations among new users and old users,this study proposes the problem of choosing opinion leader groups.Through modeling the old users in the network as multi-dimensional points,we can solve this problem by generating g-skyline groups of all points.To address the g-skyline problem,this study proposes a novel algorithm,named minimum dominance search(MDS).This algorithm first constructs a minimum dominance graph(MDG),and then search g-skyline groups based on the MDG.The experimental results show MDS is more than one magnitude faster than state-of-thearts and is effective for choosing opinion leader groups.2.To construct social relations among new users themselves,this work proposes the problem of similarity match on heterogeneous data.This problem aims to find out all qualified user pairs in the scenarios where there exists a bidirectional choice relation between two groups of users,such as online dating.This paper proposes three novel algorithms,i.e.the nested loop algorithm,the sub-match sets algorithm and the mappingfiltering-verification algorithm,to solve the problem.Comprehensive experiments are conducted to verify the effectiveness and efficiency of the three algorithms.3.To represent the social ties in a unified way,this dissertation proposes the problem of edge-based network embedding.Different from existing network embedding methods,an edge-based network embedding approach learns the embedding vectors for all the social ties and thus can benefit different analysis tasks on social ties.We propose an approach,called edge2 vec,to solve the edge-based network embedding problem.Through ingeniously combining the deep autoencoder and the skip-gram model,edge2 vec can preserve both the global and the local proximity of social ties.Experimental results demonstrate that edge2 vec can be widely used in analysis tasks about social ties.4.To analyze the directionality of different types(undirected,directed and bidirectional)of social ties,this paper proposes the concept of directionality function.The directionality function of a given social network maps each social tie to a real value which represents the directionality of this tie.Based on edge-based network embedding,this study proposes DeepDirect to learn the directionality function.It consists of two steps:learning embedding vectors in E-Step and learning directionality function in D-Step.DeepDirect learns the embedding vectors through preserving the topological proximity,utilizing labeled data and introducing directionality patterns.Through comprehensive experiments on two applications of the directionality function,this paper shows that the directionality function learned by DeepDirect are much better than those learned by baseline methods.