mvs (version 2.1.0)

MVS: Multi-View Stacking

Description

Fit a multi-view stacking model with two or more levels.

Usage

MVS(
  x,
  y,
  views,
  type = "StaPLR",
  levels = NULL,
  alphas = c(0, 1),
  nnc = c(0, 1),
  parallel = FALSE,
  seeds = NULL,
  progress = TRUE,
  relax = FALSE,
  adaptive = FALSE,
  na.action = "fail",
  na.arguments = NULL,
  ...
)
mvs(
  x,
  y,
  views,
  type = "StaPLR",
  levels = NULL,
  alphas = c(0, 1),
  nnc = c(0, 1),
  parallel = FALSE,
  seeds = NULL,
  progress = TRUE,
  relax = FALSE,
  adaptive = FALSE,
  na.action = "fail",
  na.arguments = NULL,
  ...
)

Value

An object of S3 class "MVS".

Arguments

x

input matrix of dimension nobs x nvars.

y

outcome vector of length nobs.

views

a matrix of dimension nvars x (levels - 1), where each entry is an integer describing to which view each feature corresponds.

type

a character vector of length 1 or length levels, specifying the type(s) of learner to be used at each level of MVS. Use type "StaPLR" when the desired learner(s) is/are penalized GLM(s); see StaPLR for supported families. Use type "RF" for random forests.

levels

(optional) an integer >= 2, specifying the number of levels in the MVS procedure. The default is to infer the number of levels from the supplied views argument.

alphas

a numeric vector of length levels specifying the value of the alpha parameter to use at each level.

nnc

a binary vector specifying whether to apply nonnegativity constraints or not (1/0) at each level.

parallel

whether to use foreach to fit the learners and obtain the cross-validated predictions at each level in parallel. Executes sequentially unless a parallel back-end is registered beforehand.

seeds

(optional) a vector specifying the seed to use at each level.

progress

whether to show a progress bar (only supported when parallel = FALSE).

relax

either a logical vector of length levels specifying whether model relaxation (e.g. the relaxed lasso) should be employed at each level, or a single TRUE or FALSE to enable or disable relaxing across all levels. Defaults to FALSE.

adaptive

either a logical vector of length levels specifying whether adaptive weights (e.g. the adaptive lasso) should be employed at each level, or a single TRUE or FALSE to enable or disable adaptive weights across all levels. Note that using adaptive weights is generally only sensible if alpha > 0. Defaults to FALSE.

na.action

character specifying what to do with missing values (NA). Options are "pass", "fail", "mean", "mice", and "missForest". Options "mice" and "missForest" requires the respective R package to be installed. Defaults to "fail".

na.arguments

(optional) a named list of arguments to pass to the imputation function (e.g. to mice or missForest).

...

additional arguments to pass to the learning algorithm. See e.g. StaPLR. Note that these arguments are passed to the the learner at every level of the MVS procedure.

Author

Wouter van Loon <[e-mail]>

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)

## 2-level MVS with ridge for baselearners and lasso for meta learner
views <- c(rep(1,45), rep(2,20), rep(3,20))
fit <- MVS(x=X, y=y, views=views)

## 2-level MVS with random forest for base learners and lasso for meta learner
fit <- MVS(x=X, y=y, views=views, type = c("RF", "StaPLR"))
new_X <- matrix(rnorm(2*85), nrow=2)
predict(fit, new_X)

## 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))
coefficients <- coef(fit)
predict(fit, new_X)
# }