Skip to contents

Benchmark wrapper: Pilar 10 – GAT seed-ensemble on k-NN graph

Source: R/benchmark_wosis.R
benchmark_fit_p10_gat.Rd

Builds a joint k-NN co-location graph on rbind(train, test) using (lon, lat) for adjacency and cov_cols as node features, then fits n_ensemble independent GAT regressors with different seeds, and harvests predictions at the test nodes. Predictive posterior is the seed ensemble.

Usage

benchmark_fit_p10_gat(
  train,
  test,
  cov_cols,
  k = 8L,
  hidden = 12L,
  n_heads = 2L,
  n_layers = 2L,
  epochs = 100L,
  lr = 0.03,
  n_ensemble = 10L,
  seed = 1L,
  calibrate = TRUE
)

Arguments

train, test

Data frames with soc, lon, lat, and cov_cols.

cov_cols

Character vector of covariate columns used as node features.

k

Integer; k-NN degree of the co-location graph.

hidden, n_heads, n_layers

Architecture. See gnn_fit().

epochs, lr

Training hyperparameters.

n_ensemble

Integer; number of seed-distinct fits.

seed

Base RNG seed (each member uses seed + b).

calibrate

Logical (default TRUE). When TRUE, residual noise from a representative full-data GAT fit is injected on every posterior sample (see Pilar-1 wrapper for the rationale).

Value

An edaphos_posterior with method = "ensemble", query_type = "map".