Inference on large graph

I’ve a graph with 500k nodes. Which is trained in a bipartite setting and learned the embeddings of each node. During inference when a new node comes, I want to connect the node with the neighboring nodes and perform forward pass. My question is, in this case can I just have the new node as seed node and compute GraphSAGE embeddings or do I need to compute all the embeddings for all the nodes in the network to be accurate.

You can construct the subgraph with the target node of the new node and do the neighbor sampling to create the subgrapu.

Either way is okay. The first one is less costly but might not be as accurate as the second one. A highly relevant topic is how to make transductive models inductive. There have been proposals on this in the context of getting proper embeddings for new/unseen words ( Your way of using GraphSAGE to compute an approximate embedding for the new node is definitely a reasonable thought.