cuml (version 0.2.1)

cuml_rand_proj: Random projection for dimensionality reduction.

Description

Generate a random projection matrix for dimensionality reduction, and optionally transform input data to a projection in a lower dimension space using the generated random matrix.

Usage

cuml_rand_proj(
  x,
  n_components = NULL,
  eps = 0.1,
  gaussian_method = TRUE,
  density = NULL,
  transform_input = TRUE,
  seed = 0L
)

Arguments

x

The input matrix or dataframe. Each data point should be a row and should consist of numeric values only.

n_components

Dimensionality of the target projection space. If NULL, then the parameter is deducted using the Johnson-Lindenstrauss lemma, taking into consideration the number of samples and the eps parameter. Default: NULL.

eps

Error tolerance during projection. Default: 0.1.

gaussian_method

If TRUE, then use the Gaussian random projection method. Otherwise, use the sparse random projection method. See https://en.wikipedia.org/wiki/Random_projection for details. Default: TRUE.

density

Ratio of non-zero component in the random projection matrix. If NULL, then the value is set to the minimum density as recommended by Ping Li et al.: 1 / sqrt(n_features). Default: NULL.

transform_input

Whether to project input data onto a lower dimension space using the random matrix. Default: TRUE.

seed

Seed for the pseudorandom number generator. Default: 0L.

Value

A context object containing GPU pointer to a random matrix that can be used as input to the cuml_transform() function. If transform_input is set to TRUE, then the context object will also contain a "transformed_data" attribute containing the lower dimensional projection of the input data.

Examples

# NOT RUN {
library(cuml)
library(mlbench)

data(Vehicle)
vehicle_data <- Vehicle[order(Vehicle$Class), which(names(Vehicle) != "Class")]

ctx <- cuml_rand_proj(vehicle_data, n_components = 4)

set.seed(0L)
print(kmeans(ctx$transformed_data, centers = 4, iter.max = 1000))
# }