Ranks predictors by absolute marginal correlation and keeps the top d.
This is a pre-filter, not a selector: a predictor that matters only jointly
– an interaction term, say – is marginally uncorrelated and screens out,
which is what iterated SIS was introduced to patch. Constant columns rank
last rather than producing NA.
Usage
morie_esl_sis_screening(X, y, d = NULL)
Arguments
- X
Predictors, n by p.
- y
Response.
- d
Number to keep; defaults to min(p, n - 1).
Value
List with selected (1-based column indices, strongest first),
omega, rank, d, dropped.
References
Fan, J., & Lv, J. (2008). Sure independence screening for
ultrahigh dimensional feature space. JRSS-B 70(5), 849-911.
Examples
set.seed(1)
X <- matrix(rnorm(5000), ncol = 50)
y <- 3 * X[, 8] - 2 * X[, 22] + rnorm(100, sd = 0.1)
sort(morie_esl_sis_screening(X, y, d = 5)$selected)[1:2]
#> [1] 8 15
