Convenience wrapper: build a grouped-ridge domain-adaptation model from matrices

`banded_ridge_da()` is a convenience wrapper that builds:

a `feature_sets` object for train predictors,
a test `feature_sets` object (from `X_test` or from `gamma` via `expected_features()`),
a `feature_sets_design`, and
the final `banded_ridge_da_model` spec.

Preferred name for `banded_ridge_da()`. See that function for full details.

Usage

banded_ridge_da(
  dataset,
  X_train,
  spec = NULL,
  X_test = NULL,
  gamma = NULL,
  target_builder = NULL,
  target_builder_data = NULL,
  n_test = NULL,
  drop_null = TRUE,
  renormalize = FALSE,
  block_var_test = NULL,
  ...
)

grouped_ridge_da(
  dataset,
  X_train,
  spec = NULL,
  X_test = NULL,
  gamma = NULL,
  target_builder = NULL,
  target_builder_data = NULL,
  n_test = NULL,
  drop_null = TRUE,
  renormalize = FALSE,
  block_var_test = NULL,
  ...
)

Arguments

dataset: mvpa_dataset with train_data/test_data.
X_train: Train predictor matrix (T_train x D) or a `feature_sets` object.
spec: Feature-set spec for matrix inputs, created by `blocks()` or `by_set()`. Ignored if `X_train` is already a `feature_sets`.
X_test: Optional test predictor matrix (T_test x D) or a `feature_sets` object.
gamma: Optional alignment matrix used when `X_test` is NULL. See `expected_features()`.
target_builder: Optional fold-aware callback passed through to `feature_sets_design()`. It can rebuild target predictors separately for each outer target fold and may return a `feature_sets` object, numeric matrix, or a list containing `gamma`, `X`, or `X_test`.
target_builder_data: Optional object passed through to `target_builder` as `builder_data`.
n_test: Optional target row count used when `target_builder` is provided without a fixed `X_test`.
drop_null, renormalize: Passed to `expected_features()` when using `gamma`.
block_var_test: Optional test run/block vector (length T_test).
...: Passed through to `banded_ridge_da_model()`.

Value

A model spec of class `banded_ridge_da_model`.

Details

Use this when you already have X_train (TR x features) as a single matrix and you want to declare sets via `blocks()` or `by_set()`.

Examples

set.seed(1)
dset <- gen_sample_dataset(c(3, 3, 3), nobs = 30, external_test = TRUE,
                           ntest_obs = 20)$dataset
X_enc <- matrix(rnorm(30 * 6), 30, 6)
# Soft alignment posterior (recall x encoding), rows sum to 1
gamma <- matrix(runif(20 * 30), 20, 30)
gamma <- gamma / rowSums(gamma)
recall_runs <- rep(1:2, each = 10)

ms <- grouped_ridge_da(
  dataset = dset,
  X_train = X_enc,
  spec = blocks(low = 2, mid = 2, sem = 2),
  gamma = gamma,
  block_var_test = recall_runs,
  mode = "stacked",
  lambdas = c(low = 10, mid = 10, sem = 10),
  alpha_recall = 0.2
)