SENA: Preserving Social Structure for Network Embedding

Network embedding transforms a network into a continuous feature space. Network augmentation, on the other hand, leverages this feature representation to obtain a more informative network by adding potentially plausible edges while removing noisy edges. Traditional network embedding methods are often inefficient in capturing - (i) the latent relationship when the network is sparse (the network sparsity problem), and (ii) the local and global neighborhood structure of vertices (structure preserving problem). We propose SENA, a structural embedding and network augmentation framework for social network analysis. Unlike other embedding methods which only generate vertex features, SENA generates features for both vertices and relations (edges) by minimizing a well-designed objective function composed of a loss function and a regularization. The loss function reduces the network-sparsity problem by learning from both the edges present (true edges) and absent (false edges) in the network; whereas the regularization term preserves the structural properties of the network by efficiently considering - (i) the local neighborhood of vertices and edges, and (ii) the network spectra, i.e., eigenvectors of a symmetric matrix representing the network. We compare SENA with four baseline network embedding methods, namely DeepWalk, SE, SME and TransE. We demonstrate the efficacy of SENA through a task-based evaluation setting on different real-world networks. We consider the state-of-the-art algorithms for (i) community detection, (ii) link prediction and (iii) knowledge graph query answering, and show that with SENA's representation, these algorithms achieve up to 10%, 9% and (surprisingly) 108% higher accuracy respectively compared to the best baseline embedding methods.