A Provable Framework of Learning Graph Embeddings via Summarization

Given a large graph, can we learn its node embeddings from smaller summary graph? What is the relationship between learned original graphs and their graphs? Graph representation learning plays an important role in many graph mining applications, but em-beddings of large-scale remains challenge. Recent works try to alleviate it via summarization, which typ-ically includes three steps: reducing size by combining nodes edges into supernodes superedges,learning supernode embedding on then restoring nodes. How-ever, justification behind those steps still unknown. In this work, propose GELSUMM, well-formulated framework based sum-marization, show theoretical ground learn-ing restoration with well-known approaches closed form.Through extensive experiments real-world datasets, demonstrate that our methods matching or better performance downstream tasks.This work provides analysis for summarization helps explain under-stand mechanism existing works.