gen_marginal_samples: Generate Marginal Samples of Predictive Distribution

In shrinkGPR: Scalable Gaussian Process Regression with Hierarchical Shrinkage Priors

Generate Marginal Samples of Predictive Distribution

Description

gen_marginal_samples() generates model predictions over a grid of values for one or two specified covariates, while filling in the remaining covariates either by drawing from the training data (if fixed_x is not provided) or by using a fixed values for the remaining covariates (if fixed_x is provided). The result is a set of conditional predictions that can be used to visualize the marginal effect of the selected covariates under varying input configurations.

Usage

gen_marginal_samples(
  mod,
  to_eval,
  nsamp = 200,
  fixed_x,
  n_eval_points,
  eval_range,
  display_progress = TRUE
)

Arguments

mod	A shrinkGPR or shrinkTPR object representing the fitted Gaussian/t process regression model.
to_eval	A character vector specifying the names of the covariates to evaluate. Can be one or two variables.
nsamp	Positive integer specifying the number of posterior samples to generate. Default is 200.
fixed_x	optional data frame specifying a fixed covariate configuration. If provided, this configuration is used for all nonswept covariates. If omitted, covariates are sampled from the training data.
n_eval_points	Positive integer specifying the number of evaluation points along each axis. If missing, defaults to 100 for 1D and 30 for 2D evaluations.
eval_range	optional numeric vector (1D) or list of two numeric vectors (2D) specifying the range over which to evaluate the covariates in to_eval. If omitted, the range is set to the range of the swept covariate(s) in the training data.
display_progress	logical value indicating whether to display progress bars and messages during training. The default is TRUE.

Details

This function generates conditional predictive surfaces by evaluating the fitted model across a grid of values for one or two selected covariates. For each of the nsamp draws, the remaining covariates are either held fixed (if fixed_x is provided) or filled in by sampling a single row from the training data. The selected covariates in to_eval are then varied across a regular grid defined by n_eval_points and eval_range, and model predictions are computed using calc_pred_moments.

The resulting samples represent conditional predictions across different covariate contexts, and can be used to visualize marginal effects, interaction surfaces, or predictive uncertainty.

Note that computational and memory requirements increase rapidly with grid size. In particular, for two-dimensional evaluations, the kernel matrix scales quadratically with the number of evaluation points per axis. Large values of n_eval_points may lead to high memory usage during prediction, especially when using a GPU. If memory constraints arise, consider reducing n_eval_points.

Value

A list containing posterior predictive summaries over the evaluation grid:

• mean_pred: A matrix (1D case) or array (2D case) of predicted means for each evaluation point and posterior sample.

• grid: The evaluation grid used to generate predictions. A numeric vector (1D) or a named list of two vectors (grid1, grid2) for 2D evaluations.

Examples

if (torch::torch_is_installed()) {
  # Simulate data
  set.seed(123)
  torch::torch_manual_seed(123)
  n <- 100
  x <- matrix(runif(n * 2), n, 2)
  y <- sin(2 * pi * x[, 1]) + rnorm(n, sd = 0.1)
  data <- data.frame(y = y, x1 = x[, 1], x2 = x[, 2])

  # Fit GPR model
  res <- shrinkGPR(y ~ x1 + x2, data = data)

  # Generate posterior samples
  samps <- gen_posterior_samples(res, nsamp = 1000)

  # Plot the posterior samples
  boxplot(samps$thetas)
  }

shrinkGPR documentation built on Nov. 6, 2025, 5:09 p.m.