ABCdeep.SingleInMultiple: Single-source ABC with a deep-learning (autoencoder) Step 4...

In MosaiClusteR: An Umbrella Framework for Multi-Source and Multi-Omics Clustering

Single-source ABC with a deep-learning (autoencoder) Step 4 (ABCdeep)

Description

A deep-learning variant of ABCpp.SingleInMultiple that keeps the ABC algorithm of \insertCiteAmaratunga2008MosaiClusteR intact and replaces only Step 4: instead of clustering the raw feature sub-matrix, a neural autoencoder embeds the objects into a latent space and the latent vectors are clustered into NC base clusters. The stacked label matrix is fused across sources by f.clustABC.MultiSource, exactly as in M_ABC.

Usage

ABCdeep.SingleInMultiple(
  data,
  weighting = "var",
  normalize = NULL,
  gr = c(),
  bag = TRUE,
  numsim = 1000,
  numvar = 100,
  NC = NULL,
  topk_frac = 0.2,
  latent_dim = "auto",
  hidden_units = 32L,
  ae_epochs = 80L,
  ae_lr = 0.001,
  cluster_in_latent = "ward",
  linkage = "ward.D2",
  nugget_type = "between",
  nugget_args = list(),
  min_samples = 40L,
  seed = NULL
)

Arguments

data	Numeric matrix, objects (samples) in rows, features in columns (the MosaiClusteR convention).
weighting	Feature weighting used to rank features for the top-K focus: "var", "cv", "nugget" or equal (as in ABCpp.SingleInMultiple).
normalize	Optional normalisation applied before the internal min-max scaling (see Normalization); NULL/FALSE = none.
gr	Unused; kept for interface compatibility.
bag	Logical; bootstrap the objects each iteration (Step 2).
numsim	Number of iterations R.
numvar	Kept for backward compatibility; unused.
NC	Base-cluster count (NULL = \lfloor\sqrt{N}\rfloor).
topk_frac	Fraction of the highest-weighted features the autoencoder is trained/embedded on (default 0.2). Focusing on the informative features drops the noise that would dilute a full-feature embedding on sparse signals, while keeping enough features for dense signals; the count is floored at min(G, 20) and capped at min(G, 4000).
latent_dim	Autoencoder bottleneck size: an integer, or "auto" (default) = max(NC + 2, 8), capped at the hidden width and K-1.
hidden_units	Hidden-layer width of the encoder/decoder.
ae_epochs	Training epochs for the one-time autoencoder fit.
ae_lr	Adam learning rate.
cluster_in_latent	Latent clustering (Step 4b): "ward" (default, hierarchical) or "kmeans".
linkage	Agglomeration method when cluster_in_latent = "ward".
nugget_type, nugget_args	Passed to the nugget weighting when weighting = "nugget".
min_samples	Emit a warning when nrow(data) is below this (the autoencoder is underdetermined at very small N); default 40.
seed	Optional RNG seed (controls the autoencoder init and the bootstrap).

Value

A data frame, numsim rows by one column per object; entries are base-cluster labels (0 = object not selected that iteration) – the same contract as ABCpp.SingleInMultiple.

Speed (trained once, forward pass per iteration)

The autoencoder is trained a single time on the source's top-topk_frac most-informative features; each of the numsim iterations then performs only a forward pass over the bootstrapped objects. The network is a self-contained, vectorised R autoencoder (He initialisation, ReLU, Adam, MSE) – there is no Python, TensorFlow or torch dependency, and no PCA or other fallback: a single code path.

References

\insertAllCited

See Also

M_ABCdeep, ABCpp.SingleInMultiple, f.clustABC.MultiSource

Examples

data(mosaic_toy)
lab <- ABCdeep.SingleInMultiple(mosaic_toy$List[[1]], numsim = 40, NC = 3,
                                ae_epochs = 40, seed = 1)
dim(lab)

MosaiClusteR documentation built on Oct. 25, 2026, 1:08 a.m.