Skip to contents

Diagnose dispersion mode

Source: R/auto_distribution.R
diagnose_dispersion.Rd

Determines whether dispersion should be estimated per-row, per-column, or globally by examining the coefficient of variation of per-row and per-column dispersion estimates.

Usage

diagnose_dispersion(data, model, cv_threshold = 0.5, min_mu = 1e-06)

Arguments

data

Input matrix (sparse or dense)

model

A fitted NMF model

cv_threshold

CV threshold for declaring structured dispersion. Default 0.5.

min_mu

Floor for predicted values. Default 1e-6.

Value

A list with:

mode

Recommended DispersionMode: "global", "per_row", or "per_col"

global_phi

Global dispersion estimate

row_cv

CV of per-row dispersion estimates

col_cv

CV of per-col dispersion estimates

Details

Computes moment-based dispersion estimates \(\hat{\phi} = r^2 / \mu^p\) (where \(p\) is determined by the distribution) per row and per column. If the coefficient of variation (CV) of per-row estimates exceeds cv_threshold, per-row dispersion is recommended; similarly for per-column. If both CVs are low, global dispersion suffices.

See also

auto_nmf_distribution, nmf

Examples

# \donttest{
data(aml)
model <- nmf(aml, k = 3, maxit = 10, seed = 1)
disp <- diagnose_dispersion(aml, model)
disp$mode
#> [1] "per_row"
# }