nn_method = 'FNN' is incompatible with ret_model = TRUE about uwot HOT 4 CLOSED

I will try these to find the best values for my data.

Thank you.

from uwot.

Unfortunately, FNN doesn't provide a data structure that can be used to query new data. For small datasets, you can force the use of Annoy with nn_method = "annoy".

from uwot.

Ok.
I will do this.

Another question: are there any option to pass do annoy to search for exact NN?

Thanks again.

from uwot.

No, your only option is to set search_k and n_trees to larger values.

There are some private functions in uwot, uwot:::FNN_nn (which uses FNN) and uwot:::annoy_nn (which uses Annoy), that you could use to benchmark the Annoy settings. They will return a list containing a matrix of neighbor indices and distances. The input to both these functions already needs to be a matrix, not a dataframe (another private function, uwot:::x2m can help with that if needed).

mat <- uwot:::x2m(df)
true_nn <- uwot:::FNN_nn(mat, k = 15)
approx_nn <- uwot:::annoy_nn(mat, k = 15, n_trees = 100)
# compare contents of approx_nn$idx with true_nn$idx

from uwot.