slise_initialisation_candidates: Initialise the graduated optimisation by sampling candidates

In edahelsinki/slise: Sparse Linear Subset Explanations

Initialise the graduated optimisation by sampling candidates

Description

The procedure starts with creating num_init subsets of size d. For each subset a linear model is fitted and the model that has the smallest loss is selected.

Usage

slise_initialisation_candidates(
  X,
  Y,
  epsilon,
  weight = NULL,
  beta_max = 20/epsilon^2,
  max_approx = 1.15,
  num_init = 500,
  beta_max_init = 2.5/epsilon^2,
  pca_treshold = 10,
  max_iterations = 300,
  ...
)

Arguments

X	data matrix
Y	response vector
epsilon	error tolerance
weight	weight vector (default: NULL)
beta_max	the maximum sigmoid steepness (default: 20/epsilon^2)
max_approx	the target approximation ratio (default: 1.15)
num_init	the number of initial subsets to generate (default: 500)
beta_max_init	the maximum sigmoid steepness in the initialisation
pca_treshold	the maximum number of columns without using PCA (default: 10)
max_iterations	if ncol(X) is huge, then ols is replaced with optimisation (default:300)
...	unused parameters

Details

The chance that one of these subsets contains only "clean" data is: $$ 1-(1-(1-noise_fraction)^d)^num_init $$ This means that high-dimensional data (large d) can cause issues, which is solved by using PCA (allows for sampling smaller subsets than d).

Value

list(alpha, beta)

edahelsinki/slise documentation built on Aug. 24, 2023, 11:03 p.m.