GloDyNE: Global Topology Preserving Dynamic Network Embedding

Learning low-dimensional topological representation of a network in dynamic environments is attracting much attention due to the time-evolving nature many real-world networks. The main and common objective Dynamic Network Embedding (DNE) efficiently update node embeddings while preserving topology at each time step. idea most existing DNE methods capture changes or around affected nodes (instead all nodes) accordingly embeddings. Unfortunately, this kind approximation, although can improve efficiency, cannot effectively preserve global step, not considering inactive sub-networks that receive accumulated propagated via high-order proximity. To tackle challenge, we propose novel selecting strategy diversely select representative over network, which coordinated with new incremental learning paradigm Skip-Gram based embedding approach. extensive experiments show GloDyNE, small fraction being selected, already achieve superior comparable performance w.r.t. state-of-the-art three typical downstream tasks. Particularly, GloDyNE significantly outperforms other graph reconstruction task, demonstrates its ability preservation. source code available https://github.com/houchengbin/GloDyNE