Research On Node Relevance Measure Over Signed Information Network

We are living in an interconnected world.Most of data or informational objects,individual agents are interconnected or interact with each other,forming numerous,large,interconnected,and sophisticated networks.Without loss of generality,such interconnected networks are called information networks.The networks containing the same type of objects and links are called homogeneous information networks.The networks containing different types of objects and links are called heterogeneous information networks.With the development of information networks,many real-world relationships can be represented by signed networks with positive and negative links,where positive links indicate friendships,trust and like,whereas negative links indicate foes,distrust,dislike.For example,in international relationships,there are cooperation and confrontation between countries.In the field of e-commerce,customers can give commodities good reviews or bad reviews.In the field of life science,hormones can promote or inhibit the growth of organisms.All of these can be described as signed information networks.In signed information networks,we can clearly see the cooperation or opposition between objects,which is important for our understanding and analysis of complex systems.Measuring the relevance or similarity of two nodes in information networks is important.Without loss of generality,the related degree between the same types of nodes is called similarity,while the related degree between different types of nodes is called relevance.For example,in DBLP bibliography network,we can classify authors according to the similarities between them,thus we can divide the researchers in different research domains.We can make a more accurate user profile of authors by measuring the relevance between authors and conferences.At present,the relevance measure of nodes in information networks mainly focuses on the non-signed information networks.Measuring relevance of nodes in signed information networks is still rare.In signed information networks,measuring the relevance of two nodes is challenging.Signed information networks contain both positive and negative links,and the semantics implied by them is different.Previous researches based on non-signed information networks are not suitable for signed information networks.In signed information networks,how to deal with the relationships between positive links and negative links is still an open problem.In this thesis,we study the problems of how to measure the similarity of same-typed nodes in homogeneous signed information networks and how to measure the relevance of different-typed nodes in heterogeneous signed information networks.The main contributions are summarized as follows:1.An approach named NeiSim is proposed to measure the similarities between same-typed nodes in homogeneous signed information networks.NeiSim can make full use of the semantic information implied by positive and negative links.NeiSim utilizes the different preferences for the common neighbors of source node and target node to measure the similarity between them by extending Jaccard coefficient.If there are no common neighbors,NeiSim exploits the propagation of similarity in the network based on structure balance theory.NeiSim considers both local feature and global feature.Finally,we perform extensive experimental comparisons of the proposed method against existing algorithms on real data sets(Slashdot and Epinions).Our experimental results show that NeiSim can evaluate the similarities of two nodes in homogeneous more accurately.2.A novel method named WsRel is proposed to measure the relatedness of different-typed objects in heterogeneous weighted signed social networks.Because of the rich semantic information implied by heterogeneous information networks,we use meta-path technique to capture the semantics.Particularly,WsRel first transforms a signed network into a non-signed network according to the different semantic meanings represented by positive and negative relationships.This paves the way to properly utilize negative relationships.Next,WsRel conducts random walk from the source object to the target object based on a bunch of single meta-paths separately.Finally,WsRel combines multiple meta-paths together to obtain a more comprehensive relatedness between the source object and the target object.Extensive experiments on real datasets IMDB demonstrate the superior performance of the proposed approach.