Skip to contents

Build Temporal GAM Models Across Cross-Validation Folds

Source: R/build_temporal_gam.R
build_temporal_gam.Rd

Modeling function that constructs binomial generalized additive models (GAMs) for each cross-validation fold using presence and pseudoabsence data. Each model reserves one fold as testing data and uses the remaining folds as training data. The user supplies the model formula directly using standard mgcv formula syntax, including smooth terms such as s(), te(), and ti(). Supports automatic or manual probability thresholding for converting continuous predictions to binary suitability classifications necessary for downstream analyses. The returned object follows the same structure as build_temporal_glm, build_temporal_hv, and build_temporal_rf, and is accepted directly by generate_spatiotemporal_predictions.

Usage

build_temporal_gam(partition_result, pseudoabsence_result, model_formula,
                   link = "logit", gam_params = list(method = "REML"),
                   threshold_method  = "tss",
                   output_dir = file.path(tempdir(), "GAM_Models"),
                   create_plot = TRUE, plot_palette = "Dark 2",
                   overwrite = FALSE, time_cols = NULL, verbose = TRUE)

Arguments

partition_result

List or character. Output from spatiotemporal_partition or path to an .rds file containing that output.

pseudoabsence_result

List or character. Output from generate_absences or path to an .rds file containing that output.

model_formula

Formula or character. The right-hand side of the model formula supplied as either a formula object or a character string. The response variable (presence) is always added automatically on the left-hand side, so only the right-hand side needs to be provided. Both of the following are accepted and equivalent:

  • ~ s(Var1) + s(Var2) + Var3

  • "~ s(Var1) + s(Var2) + Var3"

Standard mgcv formula syntax applies. Smooth terms are specified with s() for univariate smooths, te() for tensor product smooths of two or more variables, and ti() for tensor product interaction terms. Parametric terms can be included alongside smooth terms using +. The basis type and dimension can be controlled via arguments to s(), e.g. s(Var1, k = 5, bs = "tp"). All predictor names referenced in the formula must be present as columns in both the presence and pseudoabsence data.

Character. The link function for the binomial GAM. One of "logit" (default), "probit", "cloglog", or "cauchit". See binomial for details on each link function.

gam_params

Named list. Additional arguments passed to gam, such as method for the smoothing parameter estimation method (e.g. "REML") or select for additional shrinkage. Default is list(method = "REML").

threshold_method

Character or numeric. Method used to convert continuous predicted probabilities to binary suitability. Accepted values:

  • "prevalence": Sets threshold equal to the prevalence (proportion of presences) in the training data for that fold.

  • "tss": Selects the threshold that maximizes the True Skill Statistic (sensitivity + specificity - 1) on the training data. Default.

  • A numeric value between 0 and 1 (e.g. 0.4): Uses that value as a fixed threshold for all folds directly.

output_dir

Character. Directory to write output files including saved model objects and plots. Default is file.path(tempdir(), "GAM_Models").

create_plot

Logical. If TRUE, generates per-fold response curve plots and a combined ROC curve summary. Default is TRUE.

plot_palette

Character. Name of an HCL or RColorBrewer palette used to color folds in diagnostic plots. Accepts any HCL palette name (see hcl.pals) or, if RColorBrewer is installed, any Brewer palette name. Default is "Dark 2".

overwrite

Logical. If TRUE, overwrites existing saved model files. If FALSE, loads existing files when available. Default is FALSE.

time_cols

Character. Name of the column(s) containing year or time step values in the occurrence data. Must match time_cols used in spatiotemporal_partition. Default is NULL.

verbose

Logical. If TRUE (default), prints progress messages during processing. Includes per-fold training summaries and file-saved messages. The completion summary and metrics table are always printed regardless of this setting.

Value

A list with class "TemporalGAM" containing:

  • models: Named list of fitted gam objects, one per fold.

  • thresholds: Named numeric vector of probability thresholds used for binary classification, one per fold.

  • threshold_method: Character string recording the thresholding method used.

  • model_formula: The formula object as passed to the fitting function.

  • link: Character string recording the link function used.

  • model_vars: Character vector of predictor names extracted from the formula right-hand side.

  • fold_training_data: Named list of training data frames used to fit each fold model, retained for downstream prediction.

  • fold_test_metrics: Data frame of held-out test fold metrics per fold: Threshold, AUC, TSS, Kappa, Sensitivity, and Specificity. Also written to Fold_Test_Metrics.csv in output_dir.

  • output_dir: Path to the output directory.

  • model_type: Character string "gam", used by generate_spatiotemporal_predictions.

  • plots: Named list of recorded plot objects when create_plot = TRUE. Plots can be replayed with grDevices::replayPlot().

Details

GAMs are fit using gam from the mgcv package with family = binomial(link = link). Smooth terms default to thin plate regression splines (bs = "tp") with the basis dimension k chosen automatically by mgcv unless specified in the formula. Smoothing parameters are estimated by REML by default.

The returned object is recognized by generate_spatiotemporal_predictions, which uses the model_type field to use the correct prediction and evaluation logic.

See also

Preprocessing: spatiotemporal_partition, generate_absences

Modeling: build_temporal_glm, build_temporal_rf, build_temporal_hv, generate_spatiotemporal_predictions

External: gam, s

Examples

data(tmr_partition, package = "TemporalModelR")

data(tmr_absences,  package = "TemporalModelR")

build_temporal_gam(
  partition_result     = tmr_partition,
  pseudoabsence_result = tmr_absences,
  model_formula        = ~ s(elevation) + s(forest_cover) + s(prseas),
  threshold_method     = "tss",
  output_dir           = tempdir(),
  create_plot          = FALSE,
  time_cols            = c("year", "season"),
  verbose              = FALSE
)