Chen H.-WShuai H.-HWang S.-DYang D.-N.SHENG-DE WANG2021-09-022021-09-02202003029743https://www.scopus.com/inward/record.uri?eid=2-s2.0-85085733996&doi=10.1007%2f978-3-030-47426-3_35&partnerID=40&md5=bd9002253da80ac97aacb0552af1fe70https://scholars.lib.ntu.edu.tw/handle/123456789/581098Static network embedding has been widely studied to convert sparse structure information into a dense latent space. However, the majority of real networks are continuously evolving, and deriving the whole embedding for every snapshot is computationally intensive. To avoid recomputing the embedding over time, we explore streaming network embedding for two reasons: 1) to efficiently identify the nodes required to update the embeddings under multi-type network changes, and 2) to carefully revise the embeddings to maintain transduction over different parts of the network. Specifically, we propose a new representation learning framework, named Graph Memory Refreshing (GMR), to preserve both global types of structural information efficiently. We prove that GMR maintains the consistency of embeddings (crucial for network analysis) for isomorphic structures better than existing approaches. Experimental results demonstrate that GMR outperforms the baselines with much smaller time. ? Springer Nature Switzerland AG 2020.Data mining; Learning frameworks; Named graphs; Re-computing; Real networks; Static networks; Streaming networks; Structural information; Structure information; Embeddingsconference paper10.1007/978-3-030-47426-3_352-s2.0-85085733996