Question: Relation to NearestNeighbors.jl

[bhgomes]

I see that the interfaces are similar and that you both use the Distances.jl library for metrics. Is there any relation to this package and do you see any reason in the future to interface with them? Also, benchmarks comparing both on the same hardware would be interesting.

[dillondaudert]

This algorithm - and approximate nearest neighbor algorithms more generally - could potentially exist in NearestNeighbors.jl. The main reason I didn't pursue contributing to that package when implementing NNdescent is because the existing algorithms all output NNTrees. The output of NNdescent is a KNN graph, not a tree, so it didn't fit very naturally in that abstraction.

[dillondaudert]

As for benchmarking, adding this package to the ANN benchmarks is something I've been considering but haven't gotten around to.

[IlyaOrson]

Maybe this discussion on a common API can also be fruitful for this package.

[dillondaudert]

Thanks for the heads up. I'm in the middle of redoing the underlying implementation details in this package (in the v0.3.0-dev branch), it might be worth sharing that as part of the discussion...

[chelate]

Just a note for others, for my large and high dimensional dataset (10^6 points of dimension 30)this package is at least 10x faster than NearestNeighbors.jl and is a great help!