ARD NMF quickly finds the optimal rank for an NMF model using an exponentially variable learning rate and basic coordinate descent.
Usage
ard_nmf(
A,
n_replicates = 3,
tol = 1e-05,
maxit = 100,
verbose = 1,
L1 = 0.01,
L2 = 0,
threads = 0,
test_density = 0.05
)Arguments
- A
sparse matrix (ideally variance-stabilized) of data for genes x cells (rows x columns)
- n_replicates
number of random test sets
- tol
tolerance of the fit (1e-5 for publication quality, 1e-3 for cross-validation)
- maxit
maximum number of iterations
- verbose
no output (0/FALSE), rank-level output (1/TRUE) and step size info (2) and individual model fitting updates (3)
- L1
L1/LASSO penalty to increase sparsity of model
- L2
L2/Ridge penalty to increase angles between factors
- threads
number of threads for parallelization across CPUs, 0 = use all available threads
- test_density
fraction of values to include in the test set