Structural Representations Learning of Social Influence in Heterogeneous Social Networks

Abstract

The research of structural representations of vertexes is critical in classifying nodes, detecting communities, and predicting social links. The existing studies focus on capturing and preserving the structure of social relations, embedding the structure information into low-dimensional vector spaces. However, current approaches cannot fully consider the diversity of influence relations patterns and rich information of vertexes. To better reveal the latent patterns and preserve local and global influence structure, this paper proposes a time sequence-based semi-supervised deep synergistic method, MR-iNE (Multi-Relationship Influence Network Embedding), which mines the relationships of multi-typed of entities to capture latent influence structure, and exploits self-fusion eigenmaps reinforcement representations for vertexes in heterogeneous networks. In MR-iNE, we learn the structural representation for relationships of entities by preserving first-order, second-order, and high-order proximity influence structure and mapping them to a low-dimensional space, where fuse information from the aligned semi-supervised component that maximize the likelihood of preserving both the global and local structure of the influence relations of vertexes to overcome data sparsity. Our experiment results show the effectiveness of MR-iNE outperforms the state-of-the-art algorithms for learning rich representations on multi-label classification in real datasets from diverse domains.