randomizedRRR {SPECK}

Reduced rank reconstruction (RRR) of a matrix.

Description

Computes the rank and subsequent RRR of a m x n counts matrix. Log-normalization is first performed using the Seurat::NormalizeData() function. RRR is next performed on the normalized m x n matrix using randomized Singular Value Decomposition with the rsvd::rsvd() function. Estimated rank is selected via a construction of the standard deviations of non-centered sample principal components, which are used in a subsequent rate of change computation where each successive standard deviation value is compared to the previous to determine the rank at which the absolute value of the rate of change between consecutive values is at least 0.01 for at least two value pairs.

Usage

randomizedRRR(
  counts.matrix,
  rank.range.end = 100,
  min.consec.diff = 0.01,
  rep.consec.diff = 2,
  manual.rank = NULL,
  seed.rsvd = 1
)

Arguments

counts.matrix

A m x n counts matrix.

rank.range.end

Upper value of the rank for RRR.

min.consec.diff

Minimum difference in the rate of change between a pair of successive standard deviation estimate.

rep.consec.diff

Frequency of the minimum difference in the rate of change between a pair of successive standard deviation estimate.

manual.rank

Optional, user-specified upper value of the rank used for RRR as an alternative to automatically computed rank.

seed.rsvd

Seed specified to ensure reproducibility of the RRR.

Value

• rrr.mat - A m x n RRR matrix.

• rrr.rank - Automatically computed rank.

• component.stdev - A vector corresponding to standard deviations of non-centered sample principal components.

Examples

set.seed(10)
data.mat <- matrix(data = rbinom(n = 18400, size = 230, prob = 0.2), nrow = 80)
rrr.object <- randomizedRRR(counts.matrix = data.mat, rank.range.end = 60,
min.consec.diff = 0.01, rep.consec.diff = 2,
manual.rank = NULL, seed.rsvd = 1)
print(rrr.object$component.stdev)
print(rrr.object$rrr.rank)
dim(rrr.object$rrr.mat); str(rrr.object$rrr.mat)