rbvs: Ranking-Based Variable Selection

In rbvs: Ranking-Based Variable Selection

Description

Performs Rankings-Based Variable Selection using various measures of the dependence between the predictors and the response.

Usage

rbvs(x, y, ...)

## Default S3 method:
rbvs(x, y, m, B = 500, measure = c("pc", "dc", "lasso",
  "mcplus", "user"), fun = NULL, s.est = s.est.quotient, iterative = TRUE,
  use.residuals = TRUE, k.max, min.max.freq = 0, max.iter = 10,
  verbose = TRUE, ...)

Arguments

x	Matrix with n observations of p covariates in each row.
y	Response vector with n observations.
...	Other parameters that may be passed to fun ands.est.
m	Subsample size used in the RBVS algorithm.
B	Number of sample splits.
measure	Character with the name of the method used to measure the association between the response and the covariates. See Details below.
fun	Function used to evaluate the measure given in measure. It is required when method=="user". Must have at least three arguments: x (covariates matrix), .y (response vector), subsamples (a matrix, each row contains indices of the observations to be used); return a vector of the same length as the number of covariates in .x. See for example pearson.cor or lasso.coef.
s.est	Function used to estimate the number of important covariates based on the RBVS path. Must accept probs (a vector with probabilities) as an argument. See s.est.quotient and Details below.
iterative	Logical variable indicating the type of the procedure. If TRUE, an iterative extension of the RBVS algorithm is launched.
use.residuals	Logical. If true, the impact of the previously detected variables is removed from the response in the IRBVS procedure.
k.max	Maximum size of the subset of important variables..
min.max.freq	Positive integer. Optional parameter - the algorithm stops searching for the most frequent set when the frequencies reach this value.
max.iter	Maximum number of iterations fot the IRBVS algorithm.
verbose	Logical indicating wheter the progress of the algorithm should be reported.

Details

Currently supported measures are: Pearson correlation coefficient (measure="pc"), Distance Correlation (measure="dc"), the regression coefficients estimated via Lasso (measure="lasso"), the regression coefficients estimated via MC+ (measure="mcplus").

Value

Object of class rbvs with the following fields

measure	Character indicating type of measure used.
score	List with scores at each iteration.
subsets	A list with subset candidates at each iteration.
frequencies	A list with observed frequencies at each iteration.
ranks	Rankings evaluated (for the last iteration iterative=TRUE)
s.hat	Vector with the number of the covariates selected at each iteration.
active	Vector with the selected covariates.
timings	Vector reporting the amount of time the (I)RBVS algorithm took at each iteration.

References

R. Baranowski, P. Fryzlewicz (2015), Ranking-Based Variable Selection, in submission (http://personal.lse.ac.uk/baranows/rbvs/rbvs.pdf)).

Examples

set.seed(1)

x <- matrix(rnorm(200*1000),200,1000)
active <- 1:4
beta <- c(3,2.5,-1.7,-1)
y <- 1*rnorm(200) +x[,active]%*%beta
#RBVS algorithm
rbvs.object <- rbvs(x,y, iterative=FALSE)
rbvs.object$active
rbvs.object$subsets[[1]][[4]]
#IRBVS algorithm
rbvs.object <- rbvs(x,y)
rbvs.object$active

Example output

Iteration 1: evaluating omega
Iteration 1: evaluating rankings
Iteration 1: selecting variables
Iteration 1: 3 variables selected
[1] 1 2 3
[1] 1 2 3 4
Iteration 1: evaluating omega
Iteration 1: evaluating rankings
Iteration 1: selecting variables
Iteration 1: 3 variables selected
Iteration 1: 3 removing impact of the selected variables
Iteration 2: evaluating omega
Iteration 2: evaluating rankings
Iteration 2: selecting variables
Iteration 2: 1 variables selected
Iteration 2: 1 removing impact of the selected variables
Iteration 3: evaluating omega
Iteration 3: evaluating rankings
Iteration 3: selecting variables
Iteration 3: 0 variables selected
[1] 1 2 3 4

rbvs documentation built on May 2, 2019, 7:31 a.m.