Skip to contents

Cluster Data with a Self-Organizing Map

Source: R/som.R
get.som.codes.Rd

Replaces FlowSOM::SOM() / EmbedSOM::SOM() as the SOM clustering engine behind get.af.spectra() and get.fluor.variants(). Uses the OpenMP- accelerated batch SOM (som_train_batch_cpp()) from AutoSpectralRcpp when available, falling back to FlowSOM::SOM() (pure R, no compiled dependency) otherwise – so a plain AutoSpectral install with no AutoSpectralRcpp still works, just without the OpenMP speedup.

Usage

get.som.codes(
  data,
  som.dim,
  rlen = 10L,
  radius = NULL,
  dist = 4L,
  seed = 1337L,
  threads = 0L,
  unit.norm = FALSE
)

Arguments

data

Numeric matrix, training events x features. Must have colnames.

som.dim

Integer, side length of the square SOM grid.

rlen

Integer, number of epochs (full passes over the data). Default 10.

radius

Length-2 numeric vector, start/end neighbourhood radius. Default NULL, which derives both ends from the grid's own neighbour distances (67th percentile down to a small non-zero floor). Only used on the AutoSpectralRcpp path – FlowSOM::SOM() manages its own radius schedule internally.

dist

Integer 1:4, distance function (1 manhattan, 2 euclidean, 3 chebyshev, 4 cosine). Default 4.

seed

Integer, RNG seed for the initial codebook sample. Callers should pass asp$bird.seed.

threads

Integer, OpenMP threads for the accelerated path. Default 0 (all available cores). Ignored on the FlowSOM::SOM() fallback path, which is single-threaded.

unit.norm

Logical, default FALSE. Scale every training row to unit L2 length before training. Appropriate with dist = 4 (cosine), where the distortion-minimising code is the mean of unit event vectors rather than the mean of raw ones; without it, bright events dominate each node's code. Leaves cosine distances, and therefore node assignments, unchanged. Has no principled justification under dist = 2 and is left off by default so existing callers are unaffected. Note that the returned codes are then on the unit-normalised scale.

Value

A list with codes (matrix, SOM nodes x features), grid, and nNodes, matching the subset of FlowSOM::SOM()'s return value actually consumed elsewhere in AutoSpectral (map$codes).