I think I found the answer to my question. So if my goal is to create node embeddings using the random walk algorithm. I could potentially only store the node_ids as projected graph. Then while creating the random walk for each start node, proceed to collect all the node_ids, then make a cypher query to retrieve their properties and labels separately. By this approach, I reduce the memory footprint, but end up loosing up on time, by making one additional query per walk. Would this be a viable solution?