Repository files navigation

RcppML is an R package for fastnon-negative matrix factorization anddivisive clustering usinglarge sparse matrices. For the single-cell analysis version of functionality in RcppML, check outzdebruine/singlet.

Check out theRcppMLpkgdown site!

RcppML NMF is:

• Thefastest NMF implementation in any language for sparse and dense matrices
• Moreinterpretable than other implementations due to diagonal scaling
• Easy toregularize with an L1 penalty

Install fromCRAN or the development version from GitHub:

install.packages('RcppML')                       # install CRAN versiondevtools::install_github("zdebruine/RcppML")     # compile dev version

NOTE: RcppML is being actively developed. Please check that yourpackageVersion("RcppML") is current before raising issues.

Check out theCRAN manual.

Once installed and loaded, RcppML C++ headers defining classes can be used in C++ files for any R package using#include <RcppML.hpp>.

Sparse matrix factorization by alternating least squares:

• Non-negativity constraints
• L1 regularization
• Diagonal scaling
• Rank-1 and Rank-2 specializations (~2x faster thanirlba SVD equivalents)

Read (and cite) ourbioRXiv manuscript on NMF for single-cell experiments.

Thenmf function runs matrix factorization by alternating least squares in the formA = WDH. Theproject function updatesw orh given the other, while themse function calculates mean squared error of the factor model.

library(RcppML)A <- Matrix::rsparsematrix(1000, 100, 0.1) # sparse Matrix::dgCMatrixmodel <- RcppML::nmf(A, k = 10)h0 <- predict(model, A)evaluate(model, A) # calculate mean squared error

Divisive clustering by rank-2 spectral bipartitioning.

• 2nd SVD vector is linearly related to the difference between factors in rank-2 matrix factorization.
• Rank-2 matrix factorization (optional non-negativity constraints) for spectral bipartitioning~2x faster thanirlba SVD
• Sensitive distance-based stopping criteria similar to Newman-Girvan modularity, but orders of magnitude faster
• Stopping criteria based on minimum number of samples

Thedclust function runs divisive clustering by recursive spectral bipartitioning, while thebipartition function exposes the rank-2 NMF specialization and returns statistics of the bipartition.

library(RcppML)A <- Matrix::rsparsematrix(1000, 1000, 0.1) # sparse Matrix::dgcMatrixclusters <- dclust(A, min_dist = 0.001, min_samples = 5)cluster0 <- bipartition(A)