Create distance functions for input into precise_dist — precise_func

precise_func_fact stands for function factory, thus this function automatically creates distance functions which can be passed to the precise_dist dist_funcs parameter.

Usage

precise_func_fact(func = "rbf", params, ...)

Arguments

func: A string of the function to use. Choices include "rbf", "laplace", "minkowski", "random_forest", "kodama", "kodama_knn", "kodama_pls" and "tsne".
params: A numeric vector of values as input into the algorithm determined by the func parameter.
...: Additional named settings for func = "kodama" only: M, Tcycle, metrics, landmarks, n.cores, and seed. The KODAMA seed defaults to 1234; its execution does not alter the caller's random-number state.

Value

A list of typed functions for input into the dist_funcs argument of precise_dist.

Details

While most distance functions have no second argument, some do allowing for multiple views of a dataset tweaked by the second argument of the distance function. The following describes the params argument for each possible func choice:

For both RBF and Laplace the params argument refers to the sigma parameter of the rbfdot/laplacedot functions.
For minkowski, the params argument refers to it's power.
For random_forest, the params argument refers to mtry; the factory returns 1 - proximity as a distance matrix.
For kodama, kodama_knn and kodama_pls, the params argument refers to KODAMA::KODAMA.matrix() ncomp. The old kodama_knn/kodama_pls names are retained as aliases; current KODAMA selects the backend automatically.
For tsne, the params argument refers to perplexity.
See examples below for some reasonable defaults for the params argument for different func parameter choices.

References

Muchmore, B., Muchmore P. and Alarcón-Riquelme ME. (2018). Optimal Distance Matrix Construction with PreciseDist and PreciseGraph.

Author

Brian Muchmore

Examples

test_data <- replicate(10, rnorm(100))

# Estimate sigma for "rbf" and "laplace".
sigma_est <- kernlab::sigest(test_data, frac = 1, na.action = na.omit, scaled = FALSE)

# Build 10 values between the 0.1 and 0.9 sigma estimates.
sigma_params <- seq(sigma_est[[1]], sigma_est[[3]], length.out = 10)

# Use those values as the params for rbf or laplace factories.
rbf_funcs <- precise_func_fact(
  func = "rbf",
  params = sigma_params
)

# Build 10 integer mtry values for random_forest.
rf_params <- round(seq(2, round((ncol(test_data) * 0.5), 0), length.out = 10), 0)

# Use those values as random_forest params.
rf_funcs <- precise_func_fact(
  func = "random_forest",
  params = rf_params
)

# Build 10 Minkowski p values. p = 1 is Manhattan; p = 2 is Euclidean.
minkow_params <- seq(1, 2, length.out = 10)

# Use those values as minkowski params.
minkow_funcs <- precise_func_fact(
  func = "minkowski",
  params = minkow_params
)

# Build 10 perplexity values for tsne.
tsne_params <- seq(5, 50, length.out = 10)

# Use those values as tsne params.
tsne_funcs <- precise_func_fact(
  func = "tsne",
  params = tsne_params
)

# Combine the factories for precise_dist(dist_funcs = ...).

precise_dist_input_funcs <- rbf_funcs %>%
  append(rf_funcs) %>%
  append(minkow_funcs) %>%
  append(tsne_funcs)