Estimate Density Ratio p(x)/q(x)

Usage

densratio(
  x1,
  x2,
  method = c("uLSIF", "RuLSIF", "KLIEP"),
  sigma = "auto",
  lambda = "auto",
  alpha = 0.1,
  kernel_num = 100,
  fold = 5,
  verbose = TRUE
)

Arguments

x1: numeric vector or matrix. Data from a numerator distribution p(x).
x2: numeric vector or matrix. Data from a denominator distribution q(x).
method: "uLSIF" (default), "RuLSIF", or "KLIEP".
sigma: positive numeric vector. Search range of Gaussian kernel bandwidth.
lambda: positive numeric vector. Search range of regularization parameter for uLSIF and RuLSIF.
alpha: numeric in [0, 1]. Relative parameter for RuLSIF. Default 0.1.
kernel_num: positive integer. Number of kernels.
fold: positive integer. Numer of the folds of cross validation for KLIEP.
verbose: logical (default TRUE).

Value

densratio object that contains a function to compute estimated density ratio.

Examples

x1 <- rnorm(200, mean = 1, sd = 1/8)
x2 <- rnorm(200, mean = 1, sd = 1/2)

densratio_obj <- densratio(x1, x2)
#> ################## Start RuLSIF ##################
#> Searching optimal sigma and lambda...
#>   sigma = 0.001, lambda = 0.001, score = 15.067
#>   sigma = 0.001, lambda = 0.003, score = 0.247
#>   sigma = 0.001, lambda = 0.010, score = -0.550
#>   sigma = 0.003, lambda = 0.010, score = -1.048
#>   sigma = 0.010, lambda = 0.032, score = -1.118
#>   sigma = 0.032, lambda = 0.316, score = -1.138
#>   sigma = 0.100, lambda = 0.032, score = -1.157
#>   sigma = 0.100, lambda = 0.100, score = -1.186
#>   sigma = 0.100, lambda = 0.316, score = -1.207
#>   sigma = 0.100, lambda = 1.000, score = -1.224
#> Found optimal sigma = 0.100, lambda = 1.000.
#> Optimizing kernel weights...
#> End.
#> ################## Finished RuLSIF ###############

new_x <- seq(0, 2, by = 0.05)
estimated_density_ratio <- densratio_obj$compute_density_ratio(new_x)

plot(new_x, estimated_density_ratio, pch=19)