Shows how spatialRF::rf_interactions() works to find variable combinations (via multiplication and PCA) that improve model transferability.

engineering_features.R

#installing the development version of the package (v.1.1.1)
remotes::install_github(
  repo = "blasbenito/spatialRF", 
  ref = "development",
  force = TRUE,
  quiet = TRUE
  )
library(spatialRF)

#loading the example data
data(plant_richness_df)

#data required to execute the function
xy <- plant_richness_df[, c("x", "y")]
dependent.variable.name <- "richness_species_vascular"
predictor.variable.names <- colnames(plant_richness_df)[5:21]

#finding useful variable combinations
combinations <- rf_interactions(
  data = plant_richness_df,
  dependent.variable.name = dependent.variable.name,
  predictor.variable.names = predictor.variable.names,
  xy = xy,
  importance.threshold = 0.50, #selects predictors with importance above quantile 0.5
  cor.threshold = 0.60,        #Pearson correlation threshold to remove redundant combinations
  repetitions = 100,           #number of independent spatial folds to perform spatial cross-validation
  training.fraction = 0.75,    #fraction of records to train and evaluate models via spatial cross-validation
  seed = 1,                    #for reproducibility, results might change with different random seeds
  verbose = TRUE
)

#data frame with all the screened variable combinations
combinations$screening

#data frame with the selected variable combinations
combinations$selected

#plotting the whole thing
patchwork::wrap_plots(combinations$plot)

#fitting a model with all the variable combinations
m <- rf(
  data = combinations$data,
  dependent.variable.name = combinations$dependent.variable.name,
  predictor.variable.names = combinations$predictor.variable.names
)