Estimate Rank for NMF Models


A critical parameter in NMF algorithms is the factorization rank r. It defines the number of basis effects used to approximate the target matrix. Function nmfEstimateRank helps in choosing an optimal rank by implementing simple approaches proposed in the literature.

In the plot generated by plot.NMF.rank, each curve represents a summary measure over the range of ranks in the survey. The colours correspond to the type of data to which the measure is related: coefficient matrix, basis component matrix, best fit, or consensus matrix.


nmfEstimateRank(x, range, method = nmf.getOption("default.algorithm"), nrun = 30, 
  model = NULL, ..., verbose = FALSE, stop = FALSE)

S3 (NMF.rank)
`plot`(x, y = NULL, what = c("all", "cophenetic", "rss", "residuals", "dispersion", 
      "evar", "sparseness", "sparseness.basis", "sparseness.coef", "silhouette", "silhouette.coef", 
      "silhouette.basis", "silhouette.consensus"), na.rm = FALSE, xname = "x", yname = "y", 
      xlab = "Factorization rank", ylab = "", main = "NMF rank survey", ...)


For nmfEstimateRank a target object to be estimated, in one of the format accepted by interface nmf. For plot.NMF.rank an object of class NMF.rank as returned by function nmfEstimateRank.
a numeric vector containing the ranks of factorization to try. Note that duplicates are removed and values are sorted in increasing order. The results are notably returned in this order.
A single NMF algorithm, in one of the format accepted by the function nmf.
a numeric giving the number of run to perform for each value in range.
model specification passed to each nmf call. In particular, when x is a formula, it is passed to argument data of nmfModel to determine the target matrix -- and fixed terms.
toggle verbosity. This parameter only affects the verbosity of the outer loop over the values in range. To print verbose (resp. debug) messages from each NMF run, one can use .options='v' (resp. .options='d') that will be passed to the function nmf.
logical flag for running the estimation process with fault tolerance. When TRUE, the whole execution will stop if any error is raised. When FALSE (default), the runs that raise an error will be skipped, and the execution will carry on. The summary measures for the runs with errors are set to NA values, and a warning is thrown.
For nmfEstimateRank, these are extra parameters passed to interface nmf. Note that the same parameters are used for each value of the rank. See nmf. For plot.NMF.rank, these are extra graphical parameter passed to the standard function plot. See plot.
reference object of class NMF.rank, as returned by function nmfEstimateRank. The measures contained in y are used and plotted as a reference. It is typically used to plot results obtained from randomized data. The associated curves are drawn in red (and pink), while those from x are drawn in blue (and green).
a character vector whose elements partially match one of the following item, which correspond to the measures computed by summary on each -- multi-run -- NMF result: ‘all’, ‘cophenetic’, ‘rss’, ‘residuals’, ‘dispersion’, ‘evar’, ‘silhouette’ (and more specific *.coef, *.basis, *.consensus), ‘sparseness’ (and more specific *.coef, *.basis). It specifies which measure must be plotted (what='all' plots all the measures).
single logical that specifies if the rank for which the measures are NA values should be removed from the graph or not (default to FALSE). This is useful when plotting results which include NAs due to error during the estimation process. See argument stop for nmfEstimateRank.
legend labels for the curves corresponding to measures from x and y respectively
x-axis label
y-axis label
main title


nmfEstimateRank returns a S3 object (i.e. a list) of class NMF.rank with the following elements:

measures a data.frame containing the quality measures for each rank of factorizations in range. Each row corresponds to a measure, each column to a rank. consensus a list of consensus matrices, indexed by the rank of factorization (as a character string). fit a list of the fits, indexed by the rank of factorization (as a character string).


Note that from version 0.7, one can equivalently call the function nmf with a range of ranks.

Given a NMF algorithm and the target matrix, a common way of estimating r is to try different values, compute some quality measures of the results, and choose the best value according to this quality criteria. See Brunet et al. (2004) and Hutchins et al. (2008).

The function nmfEstimateRank allows to perform this estimation procedure. It performs multiple NMF runs for a range of rank of factorization and, for each, returns a set of quality measures together with the associated consensus matrix.

In order to avoid overfitting, it is recommended to run the same procedure on randomized data. The results on the original and the randomised data may be plotted on the same plots, using argument y.


Brunet J, Tamayo P, Golub TR and Mesirov JP (2004). "Metagenes and molecular pattern discovery using matrix factorization." _Proceedings of the National Academy of Sciences of the United States of America_, *101*(12), pp. 4164-9. ISSN 0027-8424, , .

Hutchins LN, Murphy SM, Singh P and Graber JH (2008). "Position-dependent motif characterization using non-negative matrix factorization." _Bioinformatics (Oxford, England)_, *24*(23), pp. 2684-90. ISSN 1367-4811, , .


if( !isCHECK() ){

n <- 50; r <- 3; m <- 20
V <- syntheticNMF(n, r, m)

# Use a seed that will be set before each first run
res <- nmfEstimateRank(V, seq(2,5), method='brunet', nrun=10, seed=123456)
# or equivalently
res <- nmf(V, seq(2,5), method='brunet', nrun=10, seed=123456)

# plot all the measures
# or only one: e.g. the cophenetic correlation coefficient
plot(res, 'cophenetic')

# run same estimation on randomized data
rV <- randomize(V)
rand <- nmfEstimateRank(rV, seq(2,5), method='brunet', nrun=10, seed=123456)
plot(res, rand)


plot of chunk unnamed-chunk-1