Fit an initial MRF model

Usage

mrf3_init(
  dat.list,
  ntree = 300,
  scale = TRUE,
  ytry = NULL,
  samptype = c("swor", "swr"),
  connect_list = NULL,
  filter_mode = c("auto", "none", "manual"),
  filter_method = c("mad", "variance"),
  top_n_by_type = NULL,
  top_n_manual = NULL,
  filter_verbose = TRUE,
  return_data = FALSE,
  sub_mrf = FALSE,
  sub_mrf_args = list(),
  select_connection = FALSE,
  compute_oob = FALSE,
  verbose = TRUE,
  seed = 529,
  ...
)

Arguments

dat.list: A list containing multi-omics datasets with samples in rows and features in columns.
ntree: Number of trees for fitting MRF model. Default is 300.
scale: Whether to z-standardize each feature. Default is TRUE.
ytry: Number of response variables sampled at each split. NULL delegates to the engine default.
samptype: Sampling scheme passed to forest fitting: "swor" or "swr".
connect_list: Optional pre-defined connection list. If NULL, all directed pairwise models are retained. Their modularity scores are used as soft fusion weights; set select_connection = TRUE only to request the legacy direction-selection step.
filter_mode: Feature filtering mode passed to filter_omics(): "auto", "none", or "manual".
filter_method: Feature dispersion metric passed to filter_omics(): "mad" or "variance".
top_n_by_type: Optional auto-filter overrides passed to filter_omics().
top_n_manual: Optional manual top-n configuration passed to filter_omics().
filter_verbose: Logical; whether to print filtering diagnostics.
return_data: Whether to return the data list. Default is FALSE.
sub_mrf: Logical; if TRUE, uses fit_sub_multi_rfsrc() instead of fit_multi_forest() for the final model fit. This sub-samples response and predictor features per connection, fits smaller MRFs, and averages the resulting n x n matrices. Useful when p is large. Default is FALSE.
sub_mrf_args: Named list of arguments forwarded to fit_sub_multi_rfsrc() when sub_mrf = TRUE. Common options: n_sub (number of replicates, default 15), frac_response (default 0.2), frac_predictor (default 0.2), ntree_per_sub (by default ceiling(ntree / n_sub)), min_response_for_sub (default 500; blocks smaller than this use all features), min_predictor_for_sub (default 500), and ytry (the main-model value). Unless explicitly overridden, sub-MRF fits also inherit the main forest's samptype, mtry, engine, split count, thread count, node size, and depth.
select_connection: Logical; whether to apply the legacy quality-based connection-selection step. The default is FALSE, retaining all connections for modularity-weighted fusion.
compute_oob: Logical; whether to compute OOB normalized MSE during connection selection. When TRUE, quality score = modularity - oob_nmse; when FALSE (default), quality score = modularity only, which is faster.
verbose: Logical; whether to print progress messages.
seed: Random seed.
...: Additional arguments passed to forest fitting helpers.

Value

mrf3 object

Details

mrf3_init() now performs initialization only (filtering, forest fitting, and optional connection selection). IMD weights are computed downstream via get_multi_weights() (for example inside mrf3_fit(..., run_imd = TRUE)).

When sub_mrf = TRUE, the final model fit uses a sub-sampling ensemble strategy: for each connection, features are randomly sub-sampled from both response and predictor blocks, smaller MRFs are fitted, and the n x n forest-weight and proximity matrices are averaged. This trades a small amount of matrix approximation accuracy for substantial speed gains when p is large (e.g., 6-8x faster at p = 2000). When connect_list = NULL, all directed connections are retained by default and modularity is used only as a fusion weight. Legacy direction selection is opt-in through select_connection = TRUE.