Skip to contents

An R implementation of the Uniform Manifold Approximation and Projection (UMAP) method for dimensionality reduction of McInnes et al. (2018). Also included are the supervised and metric (out-of-sample) learning extensions to the basic method. Translated from the Python implementation.


April 21 2024 As ordained by prophecy, version 0.2.2 of uwot has been released to CRAN. RSpectra is back as a main dependency and I thought I had worked out a clever scheme to avoid the failures seen in some installations with the irlba/Matrix interactions. This releases fixes the problem on all the systems I have access to (including GitHub Actions CI) but some CRAN checks remain failing. How embarrassing. That said, if you have had issues, it’s possible this new release will help you too.

April 18 2024 Version 0.2.1 of uwot has been released to CRAN. Some features to be aware of: RcppHNSW and rnndescent are now supported as optional dependencies. If you install and load them, you can use them as an alternative to RcppAnnoy in the nearest neighbor search and should be faster. Also, a new umap2 function has been added, with updated defaults compared to umap. Please see the updated and new articles on HNSW, rnndescent, working with sparse data and umap2. I consider this worthy of moving from 0.1.x to 0.2.x, but in the interests of full disclosure, on-going irlba problems has caused a CRAN check failure, so we might be onto 0.2.2 sooner than I’d like.



From github

uwot makes use of C++ code which must be compiled. You may have to carry out a few extra steps before being able to build this package:

Windows: install Rtools and ensure C:\Rtools\bin is on your path.

Mac OS X: using a custom ~/.R/Makevars may cause linking errors. This sort of thing is a potential problem on all platforms but seems to bite Mac owners more. The R for Mac OS X FAQ may be helpful here to work out what you can get away with. To be on the safe side, I would advise building uwot without a custom Makevars.




# umap2 is a version of the umap() function with better defaults
iris_umap <- umap2(iris)

# but you can still use the umap function (which most of the existing 
# documentation does)
iris_umap <- umap(iris)

# Load mnist from somewhere, e.g.
# devtools::install_github("jlmelville/snedata")
# mnist <- snedata::download_mnist()

mnist_umap <- umap(mnist, n_neighbors = 15, min_dist = 0.001, verbose = TRUE)
  cex = 0.1,
  col = grDevices::rainbow(n = length(levels(mnist$Label)))[as.integer(mnist$Label)] |>
    grDevices::adjustcolor(alpha.f = 0.1),
  main = "R uwot::umap",
  xlab = "",
  ylab = ""

# I recommend the following optional packages
# for faster or more flexible nearest neighbor search:
install.packages(c("RcppHNSW", "rnndescent"))

# Installing RcppHNSW will allow the use of the usually faster HNSW method:
mnist_umap_hnsw <- umap(mnist, n_neighbors = 15, min_dist = 0.001, 
                        nn_method = "hnsw")
# nndescent is also available
mnist_umap_nnd <- umap(mnist, n_neighbors = 15, min_dist = 0.001, 
                       nn_method = "nndescent")
# umap2 will choose HNSW by default if available
mnist_umap2 <- umap2(mnist)

Documentation For more examples see the get started doc. There are plenty of articles describing various aspects of the package.


GPLv3 or later.


If you want to cite the use of uwot, then use the output of running citation("uwot") (you can do this with any R package).

See Also