Graph-based causal discovery (Pilar 10 x Pilar 1) — gnn_causal

Runs causal_structure_learn() on a feature frame augmented with GAT node embeddings. Returns the discovered DAG restricted to the user's canonical variables; embeddings act as nuisance conditioners that absorb spatial dependence the expert DAG does not name.

Usage

gnn_causal_discovery(
  gnn_fit,
  feature_frame,
  method = c("hc", "tabu", "pc-stable"),
  whitelist = NULL,
  blacklist = NULL,
  n_emb_cols = NULL,
  bootstrap = FALSE,
  R_boot = 100L,
  seed = NULL
)

Arguments

gnn_fit: An edaphos_gnn_gat from gnn_fit().
feature_frame: Data frame of the actual variables to form the DAG over. Must have the same number of rows as the graph on which gnn_fit was trained.
method: Passed to causal_structure_learn(); default "hc".
whitelist, blacklist: Optional edge constraints (see causal_structure_learn()).
n_emb_cols: Integer; how many GAT embedding dimensions to use as conditioners. Default min(8, emb_dim) to keep the search space manageable.
bootstrap: Logical; pass to causal_structure_learn().
R_boot: Integer; bootstrap resamples.
seed: Optional RNG seed.

Value

An edaphos_causal_kg from the underlying structure-learn call, restricted to edges between variables in feature_frame.