Research On Representation Learning Algorithm Based On Nodes And Edges In Signed Networks

Social networks exist extensively in real world.Many networks in real world are bipolarly linked,i.e.entities are linked by positive and negative edges,and such kind of networks are called signed networks.For examples,in Facebook,Wechat,QQ,and other online social applications,users are connected by friend or foe relationships to form a signed network.Recently the study of signed network representation learning is the most heated signed network research among all.Network representation learning,or network embedding,aims at learning low dimensional vector representations of network entities.Since there are billions of entities in signed networks,transferring them into low-dimensional vectors helps to perform network mining tasks more efficiently.Researchers has recently developed some node representation learning algorithms for signed networks,however,the embedding accuracy of these algorithms can be further improved,the embedding quality of nodes can be further improved.Moreover,no algorithm has been proposed to learn representations for links,though they are the key components of a signed network,which makes existing works hard to perform well in link-based mining tasks.The main difficulty of utilizing feedback information is how to distinguish good pattern and bad pattern and how to adjust the embedding output;the main difficulty about edge embedding is there's hardly a method to clearly define an edge's structural property.In this paper,I propose a novel node representation learning algorithm called SignCoder,as well as an edge representation learning algorithm for signed networks,called Signed Edge Embedding(SEE).The core of SignCoder is its feedback adjustment module,which learns the underlying patterns of well-embedded node pairs and use them to adjust the representations of badly-embedded node pairs.The key idea of SEE is to extract similarity between links by using a triad-based strategy or through a adjacency link-based strategy,and embed the similarities between links into representations.Through extensive experiments on real world signed networks I have verified the superiority of SignCoder and SEE in real world networks,comparing to state-of-the-art baselines.