keep the rownames

[jwijffels]

I'm using the package here https://github.com/bnosac/ETM and here https://github.com/bnosac/word2vec for NLP purposes.
It would be great if the rownames could be kept. Is there a reason why they are dropped?

> library(uwot)
> embeddings <- matrix(runif(1000), nrow = 100, ncol = 10)
> rownames(embeddings) <- paste("word", seq_len(nrow(embeddings)))
> emb.2d <- umap(embeddings, n_components = 2, metric = "cosine", n_neighbors = 15)
> head(emb.2d)
           [,1]        [,2]
[1,]  0.4076209 -0.66765205
[2,] -0.8779236  1.09168144
[3,] -1.0821520 -1.73271100
[4,]  1.6955776 -0.17339508
[5,]  1.4383082  1.53947248
[6,] -0.5214175  0.06803199
> 
> rownames(emb.2d) <- rownames(embeddings)
> head(emb.2d)
             [,1]        [,2]
word 1  0.4076209 -0.66765205
word 2 -0.8779236  1.09168144
word 3 -1.0821520 -1.73271100
word 4  1.6955776 -0.17339508
word 5  1.4383082  1.53947248
word 6 -0.5214175  0.06803199

[jlmelville]

@jwijffels the only reason the row names were dropped was because I honestly never thought about it until you opened the issue. Thank you for reporting this.

The current master branch has the fix (it certainly works for your example). I don't know when I will be submitting a new release to CRAN, so you may want to install directly from github via the devtools or remotes packages.

[jwijffels]

Thanks for the quick fixes already. What about adding some extra check to make sure the rownames are the same, namely at

uwot/R/uwot.R

Line 2166 in 0ad9eda

nblocks <- 1

if(!all.equal(rownames(X), rownames(nn_method$idx)){
  warning("rownames of the embedding matrix not the same as the rownames from nn_method")
}

[jlmelville]

@jwijffels I can add that (or something like it), although it could be a bit of an obscure corner case: uwot::umap will either take the row names from X, or if that is not present, from the pre-computed nearest neighbor data in nn_method. Usually, you would pass either X or pre-computed nearest neighbor data, but not both. The only exception I can immediately think of is if you wanted to use pre-computed neighbors, but also initialize the embedding using PCA.

Have you had a situation where you wanted to use X and pre-computed neighbors and both had row names and the names were inconsistent? If so, are you able to provide any details about how it occurs (e.g. there is a package that is using uwot in this way and handling the row names differently inside a function so it can't be managed by the user).

[jwijffels]

I just wanted to avoid the warning message in the news file. My use cases will be mainly using uwot::umap passing on an embedding matrix with rownames and returning the model, and subsequently applying umap_transform with the model on a new set of embedding matrices containing different rownames. I think that flow is enough covered by the changes you made. Thanks again!

[jwijffels]

Let me close this. Thanks for the fixes.

[jlmelville]

Thanks @jwijffels. There are a lot of possible combinations of types of data that can be passed into uwot::umap now so I can't be 100% sure I have covered all of them in terms of getting row names out, but I am pretty confident. The NEWS caveat was just a way to cover myself by warning that if the data was inconsistent the output is implementation dependent (which at this point could change if I found a hole that needed plugging).

[jwijffels]

There are a lot of possible combinations of types of data that can be passed into uwot::umap

I saw that when inspecting the code :)