Find optimal directional connections from fitted RF models

Scores each directional RF model using modularity of the diagonal-adjusted forest weight matrix (and optionally OOB normalized MSE). Two-step selection: (1) drop the bottom drop_bottom_q fraction of connections by rank-sum of quality metrics; (2) select_one_per_pair keeps at most one direction per omics pair (lowest rank-sum).

Usage

find_connection(
  mod.list,
  return_score = FALSE,
  drop_bottom_q = 0.2,
  select_one_per_pair = TRUE,
  compute_oob = FALSE,
  seed = 529L,
  ...
)

Arguments

mod.list

A named list of fitted RF models. Model names must follow response_predictor convention.

return_score

Logical; whether to return the directional score matrix.

drop_bottom_q

Proportion of models to drop based on quality rank-sum. Must be in [0, 1). Default is 0.2.

select_one_per_pair

Logical; whether to keep at most one direction per omics pair among quality-filtered models.

compute_oob

Logical; whether to compute OOB normalized MSE in addition to modularity. When TRUE, quality_score = modularity - oob_nmse and rank-sum uses both ranks; when FALSE, quality_score = modularity and rank_sum = rank(-modularity). Default is FALSE.

seed

Random seed used by Louvain modularity scoring.

...

Additional arguments (currently ignored).

Value

A character vector of selected model names (response_predictor). If return_score = TRUE, returns a list with model_connection, connect_list, score, and model-level quality metrics.