I have a graph containing people connecting to movies they like, which I can use fastrp to generate embeddings. When I have new people nodes coming in, I want to generate embeddings just for these new people. Do I still have to run fastRP on the whole graph? I'm not clear how to use fastRP as an inductive model suitable for industry prediction as mentioned in the docs.
Thanks
Hi @xtfocus,
Sorry for the late reply. At this point you cannot run FastRP incrementally to only generate embeddings for newly added nodes.
However, this is not really a problem for most people. You can use a randomSeed in the algorithm configuration to make sure that nodes always are assigned the same start vector (seeding is based on node IDs from database). This means that if there's just a small change to the graph, embeddings after running FastRP will likely be very similar. If some are slightly different, that's probably also for good reason since the graph has been updated. Furthermore, FastRP is a very fast algorithm - linear with small constant multiplier - with a performant implementation in GDS, so for most people it's not really a problem to rerun the algorithm on the entire graph.
So while you technically cannot run the algorithm incrementally, you can probably get close to (if not exactly) the behavior you want anyway.
Hope this is helpful. Good luck,
Adam