A Provable Framework of Learning Graph Embeddings via Summarization