Structure-Enhanced Graph Representation Learning for Link Prediction in Signed Networks

Link prediction in signed networks has attracted widespread attention from researchers recently. Existing studies usually learn a representation vector for each node, which is used for link prediction tasks, by aggregating the features of neighbour nodes in the network. However, how to incorporate structural features, e.g., community structure and degree distribution, into graph representation learning remains a difficult challenge. To this end, we propose a novel Structure-enhanced Graph Representation Learning method called SGRL for link prediction in signed networks, which enables the incorporation of structural features into a unified representation. Specifically, the feature of community structure is described by introducing two latent variables to submit to Bernoulli distribution and Gaussian distribution. Moreover, the degree distribution of each node is described by a hidden variable that submits to the Dirichlet distribution by using the community feature as the parameter. Finally, the unified representation obtained from the Dirichlet distribution is further employed for the link prediction based on similarity computation. The effectiveness of the SGRL is demonstrated using benchmark datasets against the state-of-the-art methods in terms of signed link prediction, ablation study, and robustness analysis.