Network Embedding With Completely-Imbalanced Labels

Network embedding, aiming to project a network into low-dimensional space, is increasingly becoming focus of research. Semi-supervised embedding takes advantage labeled data, and has shown promising performance. However, existing semi-supervised methods would get unappealing results in the completely-imbalanced label setting where some classes have no nodes at all. To alleviate this, we propose two novel methods. The first one shallow method named RSDNE. Specifically, benefit from labels, RSDNE guarantees both intra-class similarity inter-class dissimilarity an approximate way. other RECT which new class graph neural networks. Different RSDNE, explores class-semantic knowledge. This enables handle networks with node features multi-label setting. Experimental on several real-world datasets demonstrate superiority proposed