Fit an ensemble of sub-sampled MRF models for a single connection

When the number of response and/or predictor features is large, fitting a single MRF with all features is computationally expensive. fit_sub_mrf() addresses this by repeatedly sub-sampling features from both the response and predictor blocks of a single connection, fitting a smaller MRF each time, and averaging the resulting forest-weight and proximity matrices.

Usage

fit_sub_mrf(
  X,
  Y,
  n_sub = 15L,
  frac_response = 0.2,
  frac_predictor = 0.2,
  ntree_per_sub = 25L,
  mtry = NULL,
  ytry = NULL,
  min_response = 20L,
  min_predictor = 30L,
  enhanced = FALSE,
  compute_imd = FALSE,
  imd_args = list(),
  seed = 529L,
  parallel = FALSE,
  cores = NULL,
  verbose = TRUE,
  ...
)

Arguments

X: Predictor data frame (samples x features) — a single omics block.
Y: Response data frame (samples x features) — a single omics block.
n_sub: Integer; number of sub-MRF replicates to fit.
frac_response: Fraction of response columns to sample per replicate (default 0.2). Convergence analysis shows symmetric low fractions (0.1–0.3 for both response and predictor) yield the best approximation of the full forest-weight matrix.
frac_predictor: Fraction of predictor columns to sample per replicate (default 0.2). Avoid setting this to 1.0 when frac_response is small, as the asymmetric case leads to poor convergence (each sub-model overfits to its response subset).
ntree_per_sub: Number of trees per sub-MRF (default 25 for direct calls; the workflow inherits the main tree budget by using ceiling(ntree / n_sub) unless explicitly overridden).
mtry: Number of candidate X variables per split. Passed through to fit_forest().
ytry: Number of candidate Y variables per split. NULL uses the multiRF default, ceiling(qy / 3).
min_response: Minimum number of response columns per sub-MRF.
min_predictor: Minimum number of predictor columns per sub-MRF.
enhanced: Logical; if TRUE, compute soft enhanced proximity using full-data sample embeddings within each sub-model. The result is stored as $enhanced_prox in the output so that mrf3_cl_prox() can use it directly without re-traversing trees. Default FALSE because the extra computation is non-trivial.
compute_imd: Logical; whether to pre-compute IMD weights across sub-models.
imd_args: Named list of arguments passed to get_imp_forest() when compute_imd = TRUE.
seed: Base random seed.
parallel: Logical; if TRUE, fit replicates in fresh PSOCK worker processes.
cores: Number of cores when parallel = TRUE; NULL (default) uses parallel::detectCores() - 1.
verbose: Logical; print progress messages.
...: Additional arguments forwarded to fit_forest().

Value

An rfsrc-like list with:

forest.wt: Averaged n x n forest-weight matrix (all samples).
forest.wt.oob: Averaged n x n OOB forest-weight matrix.
proximity: Averaged n x n proximity matrix.
xvar: Full predictor data (all columns of X).
yvar: Full response data (all columns of Y).
ntree: Total tree count across all sub-MRFs.
sub_mrf_info: List with coverage and timing metadata.

Details

Both forest.wt and proximity are n x n matrices that do NOT depend on feature dimension. Averaging across sub-models with different feature subsets produces a smoothed estimate.

forest.wt.oob is computed from each sub-model's membership and inbag matrices: for sample i, only trees where i is out-of-bag contribute to its weight row. This provides a regularized version suitable for connection selection via find_connection().