Research On Representation Learning Methods For Signed Network Analysis

Signed network is a form of data representation in the era of big data,which consists of positive and negative links between nodes.Tasks in signed networks have a very wide range of real-life applications,such as link prediction,community detection,node classification,etc.However,the nodes and links of the signed network require high-dimensional vector representation,which leads to complex model computation.Representation learning is a very effective tool for learning node embeddings,embedding network vertices into lowdimensional vector spaces by preserving topology,vertex content,and other auxiliary information in the network.Therefore,the representation learning method is used to effectively solve the problem of high-dimensional representation of nodes in signed networks.This paper analyzes two main tasks in signed networks,namely link prediction and community detection,and explores the techniques and applications of representation learning methods in these two tasks.The main challenges faced by the above work include: 1)Link-oriented prediction: how to integrate the community structure and node degree distribution information in the signed network into the node representation,enrich the node representation information,and improve the link prediction ability of the model? 2)Community-oriented detection: how to learn the vector representation of nodes and communities,explore the differences and connections between communities and communities,and between communities and nodes,so as to assign nodes to corresponding communities?Based on the above problems,this paper proposes two novel signed network representation learning algorithms:(1)A structure-enhanced signed network link prediction method SGRL is proposed.The current signed network link prediction methods only focus on constructing a graph neural network with balanced theory,carrying out message passing,and obtaining the vector representation of nodes.But this method only captures the topology of the signed network and the information of some neighbor nodes.However,a complex signed network has a lot of structural information,and each structural information can affect the information representation of nodes and thus affect the signed network link prediction task.Aiming at this problem,this paper proposes the SGRL method,which explores the structural information in the signed network,such as the community structure and the degree distribution information of nodes,so as to enhance the information representation ability of nodes and improve the ability of model link prediction.First,the first step of the model uses Bernoulli distribution and Gaussian distribution to describe the community structure in the signed network.Bernoulli distribution describes whether a node belongs to a certain community,and Gaussian distribution describes the strength of a node belonging to a certain community.Secondly,in the second step of the model,the Dirichlet distribution is introduced to describe the degree distribution information of the nodes.The Dirichlet distribution belongs to the power-law distribution.Therefore,the degree distribution of the nodes can be well described.Finally,node representations sampled from the Dirichlet distribution are used for the link prediction task in signed networks.The experimental results show that the community structure and the degree distribution information of the nodes are well integrated into the node representation,which improves the accuracy of the model link prediction.(2)A community detection algorithm based on the analysis of the relationship between communities and nodes is proposed.At present,there are few research on community detection in signed networks.These studies are mainly based on spectral methods,which use feature roots to explore communities.However,spectral methods have high computational complexity when dealing with large graph data,and deep learning methods,especially graph neural network methods,are more conducive to complex network community detection tasks.Aiming at this problem,this paper proposes a method based on graph neural network.The main method optimizes node representation by simulating the k-means process,and obtains the connection and difference between communities and between communities and nodes.First,the representation vector of the community is randomly initialized,and the node representation is obtained by using the graph neural network model based on the signed network,so as to determine the probability of the node belonging to the community according to the similarity between the community representation vector and the node representation,that is,the assignment matrix.Then,the parameters in the graph neural network are optimized by using the loss function constructed according to the nature of the community structure of the signed network,the node representation is updated,and the final assignment matrix is obtained to complete the community detection task.The experimental results show that the model finds the connections and differences between communities nodes,thereby improving the ability of the model to discover communities.