optimize_hinge_general: Optimize a linear classifier with possibly negative...

In aaronjfisher/mcr: Computes model class reliance, i.e., the most and least that models from a class may rely on variables of interest

Description

The optimization uses simulated annealing (SA), with a gradient-based search at each step. (Used in our paper, but not the primary focus of this package; the default option of ignore_below_zero=TRUE may cause problems in uses of this function beyond the scope of MCR, outside of this package.)

Usage

optimize_hinge_general(y, X, reg_matrix = diag(p), n = length(y),
  case.weights = rep(1/n, n), reg_threshold, K = 1000, p = dim(X)[2],
  start = NULL, constr_tol = 10^-4, sparse_warning = TRUE,
  maxit_SA = 1000, extra_step_NM = TRUE, short_cg_lim = 15,
  long_cg_lim = 500, ignore_below_zero = TRUE)

Arguments

y	outcome vector with elements -1 or 1.
X	covariate matrix (n x p), which should not contain an interecept or constant column.
reg_matrix	Matrix R, where w'Rw is the penalty of a coefficient vector w.
n	sample size
case.weights	vector of numeric multipliers (possibly negative) for each observation, when computing loss summation.
reg_threshold	Value for w'Rw above which the penalty is applied
K	penalty multiplier for w'Rw
p	covariate dimension
start	starting value for coefficient vector
constr_tol	buffer for regularization, after which a penalty is applied
sparse_warning	warn user if intercept column is detected
maxit_SA	number of iterations for SA steps
extra_step_NM	whether to follow the SA search with an additional Nelder-Mead search
short_cg_lim	tuning parameter used to set initial value of the search
long_cg_lim	how many gradient-based steps to take at each SA iteration
ignore_below_zero	stop SA search if a value is discovered below zero. This is irrelevant if weights are positive, and is useful within the MCR binary search, but may cause problems for other applications of this function beyond the computation of MCR.

Value

a linear coefficient vector

aaronjfisher/mcr documentation built on Jan. 2, 2020, 4:38 p.m.