RoSANE: Robust and Scalable Attributed Network Embedding for Sparse Networks

Attributed networks can better describe the real-world complex systems where the interaction or relationship between entities can be represented as networks and the auxiliary information can be represented as node attributes. Attributed Network Embedding (ANE) is attracting much attention. It utilizes network topology and node attributes to jointly learn enhanced low-dimensional node embeddings so as to facilitate various downstream inference tasks. However, the existing ANE methods cannot effectively embed attributed sparse networks which are important real-world scenarios, and/or are not scalable to large-scale networks. To tackle these challenges, we first integrate network topology and node attributes to reconstruct an enriched denser network, and then learn node embeddings upon the denser network. In above two steps, the techniques such as Ball-tree K-Nearest Neighbors and random walks based Skip-Gram model are adopted to guarantee the scalability, which is demonstrated via theoretical complexity analysis. The extensive empirical studies show the effectiveness and effciency of the proposed method, as well as its robustness to different
networks or the same network with different sparsities.