lmcinnes
Seeding index with a precalculated KNN graph
Hi,
Sorry if this sounds completely outlandish. This may be somewhat related to #79.
The idea is to provide a pre-calculated graph to the index. And at this 'index seeding' step, a different distance metric than the one used to calculate the seed graph may be used. Thereafter, this graph can be updated with further data batches or/and queried. Do you think the current implementation of pynndescent makes this feasible? If yes, what would be assumptions about the seed graph that one might need to take care of?
That's not a bad idea if people have a KNN graph already. So right now we "seed" a knn-graph based on results from an rp-forest. It would not be hard, I think to add an alternative to allow a user supplied knn-graph. Do you have an interest in attempting a pull request?
