kliep: Kullback-Leibler importance estimation procedure

In densityratio: Distribution Comparison Through Density Ratio Estimation

Kullback-Leibler importance estimation procedure

Description

Kullback-Leibler importance estimation procedure

Usage

kliep(
  df_numerator,
  df_denominator,
  scale = "numerator",
  nsigma = 10,
  sigma_quantile = NULL,
  sigma = NULL,
  ncenters = 200,
  centers = NULL,
  cv = TRUE,
  nfold = 5,
  epsilon = NULL,
  maxit = 5000,
  progressbar = TRUE
)

Arguments

df_numerator	data.frame with exclusively numeric variables with the numerator samples
df_denominator	data.frame with exclusively numeric variables with the denominator samples (must have the same variables as df_denominator)
scale	"numerator", "denominator", or NULL, indicating whether to standardize each numeric variable according to the numerator means and standard deviations, the denominator means and standard deviations, or apply no standardization at all.
nsigma	Integer indicating the number of sigma values (bandwidth parameter of the Gaussian kernel gram matrix) to use in cross-validation.
sigma_quantile	NULL or numeric vector with probabilities to calculate the quantiles of the distance matrix to obtain sigma values. If NULL, nsigma values between 0.25 and 0.75 are used.
sigma	NULL or a scalar value to determine the bandwidth of the Gaussian kernel gram matrix. If NULL, nsigma values between 0.25 and 0.75 are used.
ncenters	Maximum number of Gaussian centers in the kernel gram matrix. Defaults to all numerator samples.
centers	Option to specify the Gaussian samples manually.
cv	Logical indicating whether or not to do cross-validation
nfold	Number of cross-validation folds used in order to calculate the optimal sigma value (default is 5-fold cv).
epsilon	Numeric scalar or vector with the learning rate for the gradient-ascent procedure. If a vector, all values are used as the learning rate. By default, 10^{1:-5} is used.
maxit	Maximum number of iterations for the optimization scheme.
progressbar	Logical indicating whether or not to display a progressbar.

Value

kliep-object, containing all information to calculate the density ratio using optimal sigma and optimal weights.

References

Sugiyama, M., Suzuki, T., Nakajima, S., Kashima, H., Von Bünau, P., & Kawanabe, M. (2008). Direct importance estimation for covariate shift adaptation. Annals of the Institute of Statistical Mathematics 60, 699-746. Doi: https://doi.org/10.1007/s10463-008-0197-x.

Examples

set.seed(123)
# Fit model
dr <- kliep(numerator_small, denominator_small)
# Inspect model object
dr
# Obtain summary of model object
summary(dr)
# Plot model object
plot(dr)
# Plot density ratio for each variable individually
plot_univariate(dr)
# Plot density ratio for each pair of variables
plot_bivariate(dr)
# Predict density ratio and inspect first 6 predictions
head(predict(dr))
# Fit model with custom parameters
kliep(numerator_small, denominator_small,
      nsigma = 1, ncenters = 100, nfold = 10,
      epsilon = 10^{2:-5}, maxit = 500)

densityratio documentation built on June 8, 2025, 11:17 a.m.