mvs (version 2.0.0)

MRM: Minority Report Measure

Description

Calculate the Minority Report Measure (MRM) for each view in a (hierarchical) multi-view stacking model.

Usage

MRM(fit, constant, level = 2, a = 0, b = 1, cvlambda = "lambda.min")
mrm(fit, constant, level = 2, a = 0, b = 1, cvlambda = "lambda.min")

Value

A numeric vector of a length equal to the number of views at the specified level, containing the values of the MRM for each view.

Arguments

fit

an object of class MVS.

constant

the value at which to keep the predictions of the other views constant. The recommended value is the mean of the outcome variable.

level

the level at which to calculate the MRM. In a 3-level MVS model, level = 2 (the default) is generally the level for which one would want to calculate the MRM. Note that calculating the MRM for level = 1 (the feature level) is possible, but generally not sensible except under specific conditions.

a

the start of the interval over which to calculate the MRM. Defaults to 0.

b

the end of the interval over which to calculate the MRM. Defaults to 1.

cvlambda

denotes which values of the penalty parameters to use for calculating predictions. This corresponds to the defaults used during model fitting.

Author

Wouter van Loon <[e-mail]>

Details

The Minority Report Measure (MRM) considers the view-specific sub-models at a given level of the hierarchy as members of a committee making predictions of the outcome variable. For each view, the MRM quantifies how much the final prediction of the stacked model changes if the prediction of the corresponding sub-model changes from a to b, while keeping the predictions corresponding to the other views constant at constant. For more information about the MRM see <doi:10.3389/fnins.2022.830630>.

Examples

# \donttest{ 
set.seed(012)
n <- 1000
X <- matrix(rnorm(8500), nrow=n, ncol=85)
beta <- c(rep(10, 55), rep(0, 30)) * ((rbinom(85, 1, 0.5)*2)-1)
eta <- X %*% beta
p <- 1 /(1 + exp(-eta))
y <- rbinom(n, 1, p)

## 3-level MVS
bottom_level <- c(rep(1:3, each=15), rep(4:5, each=10), rep(6:9, each=5))
top_level <- c(rep(1,45), rep(2,20), rep(3,20))
views <- cbind(bottom_level, top_level)
fit <- MVS(x=X, y=y, views=views, levels=3, alphas=c(0,1,1), nnc=c(0,1,1))
MRM(fit, constant=mean(y))
# }