Train a CLR + MLP Classifier — train_clr

Fits a dense CoDA baseline: a centered log-ratio transform followed by a multilayer perceptron. The default backend uses torch when it is available, and otherwise falls back to `nnet::nnet()` or logistic regression for environments without torch.

Usage

train_clr_mlp(
  X_train,
  y_train,
  epochs = 30L,
  lr = 0.001,
  batch_size = 256L,
  class_weight = TRUE,
  device = NULL,
  verbose = TRUE,
  seed = 42L,
  positive = NULL,
  input_scale = c("log", "raw"),
  backend = c("auto", "torch", "nnet", "glm"),
  hidden_units = c(152L, 136L),
  dropout = 0.3
)

Arguments

X_train: Numeric training matrix with samples in rows and features in columns.
y_train: Binary outcome vector.
epochs: Number of training epochs. Defaults to `30`.
lr: Learning rate. Defaults to `0.001`.
batch_size: Mini-batch size. Defaults to `256`.
class_weight: Logical; if `TRUE`, use inverse-frequency weighting.
device: One of `"cpu"`, `"cuda"`, or `NULL`/`"auto"`.
verbose: Logical; print training progress. Defaults to `TRUE`.
seed: Integer random seed. Defaults to `42`.
positive: Optional positive-class label for non-numeric outcomes.
input_scale: Either `"log"` (default) or `"raw"`.
backend: One of `"auto"`, `"torch"`, `"nnet"`, or `"glm"`.
hidden_units: Integer vector with the two hidden-layer sizes used by the torch backend. Defaults to `c(152, 136)`.
dropout: Dropout rate used by the torch backend. Defaults to `0.3`.

Value

An object of class `"omicselector_clr_mlp"` suitable for [predict_clr_mlp()].

References

Aitchison J. (1986). The Statistical Analysis of Compositional Data. Chapman and Hall.

Egozcue JJ, Pawlowsky-Glahn V, Mateu-Figueras G, Barcelo-Vidal C. (2003). Isometric logratio transformations for compositional data analysis. Mathematical Geology, 35(3), 279-300.

Gloor GB, Macklaim JM, Pawlowsky-Glahn V, Egozcue JJ. (2017). Microbiome datasets are compositional: and this is not optional. Frontiers in Microbiology, 8, 2224.

Quinn TP, Erb I, Richardson MF, Crowley TM. (2020). Understanding sequencing data as compositions: an outlook and review. Bioinformatics, 36(16), 4424-4432.

Examples

X_train <- matrix(rnorm(14 * 20), nrow = 20)
y_train <- factor(rep(c("control", "case"), each = 10))
fit <- train_clr_mlp(X_train, y_train, backend = "glm", verbose = FALSE)
#> Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred
fit$backend
#> [1] "glm"