What is your opinion on network representation learning? Can it be used in Dgraph? For instance, do you think it is possible to use node embeddings as indices for faster retrieval?