MRF variable selection

Usage

mrf3_vs(
  mod,
  dat.list = NULL,
  method = "filter",
  signal = c("shared", "specific", "all"),
  se = 1,
  c1 = "normal",
  c2 = "normal",
  level = 0.05,
  tscore = FALSE,
  use_distribution = TRUE,
  re_weights = FALSE,
  re_fit = TRUE,
  ntree = NULL,
  scale = FALSE,
  k = 3,
  tau_grid = seq(0.8, 3.1, by = 0.1),
  tol = 0.01,
  perm_B = 20L,
  perm_quantile = 0.95,
  iter = 1000,
  eps = 1e-05,
  normalized = FALSE,
  select = "ALL",
  ...
)

Arguments

mod: A fitted mrf3_fit object, or an mrf3-like object that already contains raw forest IMD in mod$imd.
dat.list: A named list of omics matrices used for feature selection and optional refitting.
method: Feature-selection rule. "filter" adaptively tunes the cutoff tau * sd(IMD) using OOB error: for the shared signal the cross-omics forests are refitted at each candidate cutoff and scored by OOB normalized prediction error across predictor and response coordinates, and the cutoff is the first stable plateau of that error; for the specific signal there is no second block to predict, so by default the cutoff is the perm_quantile quantile of a permutation null: the residual forest is refitted perm_B times on column-wise permuted residuals (no between-variable structure, same marginals) and the pooled IMD of those refits defines the importance reachable without structure (an OOB search is not used for the specific signal because the OOB error of a residual block barely changes with the cutoff). The specific branch requires the residual matrices kept by mrf3_fit(); "mixture" fits a point-mass/two-component model; and "test" (alias "transformation") selects by a forest-level t-score: each feature's forest IMD is standardized against the mean forest IMD of all features in the block, using the feature's across-tree standard error, and features in the upper tail of a Student-t reference with ntree - 1 degrees of freedom (p < level) are kept; see Details. "thres" applies a fixed se * sd(IMD) cutoff.
signal: Which signal component to select on: "shared" (cross-modal), "specific" (per-block residual), or "all" (the union of both).
se: The fixed value of tau used by method = "thres".
c1: Distribution family for the lower-IMD (noise) component in mixture mode: "normal" or "truncn".
c2: Distribution family for the higher-IMD component in mixture mode: "normal" or "gamma".
level: One-sided significance level for the transformation, or the maximum posterior noise probability for mixture selection. A single numeric value is applied to every block; a named vector/list supplies block-specific values. "auto" retains the optional Bayesian-FDR rule in mixture mode.
tscore: Deprecated compatibility switch. Use method = "transformation" instead.
use_distribution: Deprecated compatibility argument.
re_weights: Logical; whether selected IMD values are supplied as variable-sampling weights during the final refit.
re_fit: Logical; whether to refit forests after feature selection.
ntree: Number of trees used in the optional final refit. NULL reuses the fitted object's tree count, falling back to 300.
scale: Logical; whether to standardize selected data before refitting.
k: Number of repeated forest fits at each candidate filtering cutoff.
tau_grid: Candidate multipliers of sd(IMD) evaluated by method = "filter". Default seq(0.8, 3.1, by = 0.1).
tol: Tolerable adjacent change in mean OOB normalized MSE used to choose the filtering cutoff.
perm_B: Number of permutation refits per residual block used by method = "filter" for the specific signal.
perm_quantile: Quantile of the pooled permutation-null IMD used as the specific-signal cutoff.
iter: Maximum EM iterations in mixture mode.
eps: EM convergence tolerance in mixture mode.
normalized: Logical; whether to L2-normalize the retained weights after selection. These methods are always evaluated on raw IMD; therefore the default is FALSE.
select: Data block(s) to which selection is applied; default "ALL".
...: Additional arguments passed to forest fitting.

Value

An object inheriting from mrf3 and vs. For signal = "all", dat.list and imd contain the union selection and the two component results are retained in signal_results.

Details

For method = "transformation", every connected forest supplies a per-tree IMD matrix (features x trees) per block. Feature v is standardized as t_v = (M_v - mu) / SE(M_v), where M_v is the feature's forest IMD (its per-tree IMD averaged over the B trees), mu is the mean forest IMD over all features of the block, and SE(M_v) is the feature's across-tree standard error (sd(per-tree IMD) / sqrt(B)). Features with an upper-tail p = P(T_{B-1} > t_v) < level are selected, i.e. features whose forest IMD lies significantly above the block-wide mean. When several connected forests cover the same block, each forest is evaluated separately and a feature is retained by strict majority vote across forests; the reported t-scores and p-values are averaged. Because per-tree IMDs within a forest are correlated (trees share the training data) and mu is treated as a fixed reference, the p-values are screening scores rather than exact test levels.