Representation learning for graph data has gained a lot ofattention in recent years. However, state-of-the-art research is focusedmostly on node embeddings, with little effort dedicated to the closelyrelated task of computing subgraph embeddings. Subgraph embeddingshave many applications, such as community detection, cascade predic-tion, and question answering. In this work, we propose a subgraph to sub-graph proximity measure as a building block for a subgraph embeddingframework. Experiments on real-world datasets show that our approach,SubRank, outperforms state-of-the-art methods on several importantdata mining tasks.