Open davidechicco opened 5 months ago
Unfortunately, there are no trustworthiness and continuity functions in uwot
. I wrote some evaluation functions at https://github.com/jlmelville/quadra but I am not sure how valuable they are.
Personally, I find that evaluating the nearest neighbor preservation at a handful of nearest neighbor values (e.g. 15, 50, 150) works well enough for me. If you are prepared to install https://github.com/jlmelville/rnndescent then it's not too hard to evaluate the overlap at whatever value of n_neighbors
you used e.g.:
library(uwot)
library(rnndescent)
n_neighbors <- 10
res <- umap(iris, ret_nn = TRUE, n_neighbors = n_neighbors)
high_dim_nn <- res$nn$euclidean
# brute_force_knn and neighbor_overlap both come from rnndescent
low_dim_nn <-
brute_force_knn(res$embedding, k = n_neighbors, n_threads = 6)
neighbor_overlap(high_dim_nn, low_dim_nn)
Hi Thanks for having developed and released the
uwot
R package. Is there a function that can express trustworthiness and continuity of the UMAP results?Best regards,
-- Davide Chicco