penalize: Penalize a latent variable model

In bozenne/lava.penalty: Penalized Latent Variable Models

Description

Add a penalty term to a latent variable model

Usage

penalize(x, ...)

penalize(x, ...) <- value

## S3 method for class 'lvm'
penalize(x, value = NULL, ...)

## S3 replacement method for class 'lvm'
penalize(x, group, add = TRUE, reduce = FALSE,
  Vlasso = TRUE, Vridge = TRUE, Vgroup = TRUE, intercept = FALSE,
  regression = TRUE, variance = FALSE, covariance = FALSE,
  latent = FALSE, ...) <- value

penalizeNuclear(x, ...)

penalizeNuclear(x, ...) <- value

## S3 method for class 'lvm'
penalizeNuclear(x, value = NULL, ...)

## S3 replacement method for class 'lvm'
penalizeNuclear(x, coords = NULL, lambdaN = NULL,
  ...) <- value

Arguments

x	lvm-object
value	the name of the link to be penalized
group	the groups defining the group lasso penalty
add	should value be added to the existing penalty term ? Otherwise it will overwrite it.
reduce	should for each regression the penalised link be aggregated into a linear predictor.
Vlasso	the matrix defining lasso penalties. Otherwise must be logical to indicate whether to add lasso penalties.
Vridge	the matrix defining ridge penalties. Otherwise must be logical to indicate whether to add ridge penalties.
Vgroup	the matrix defining group lasso penalties. Otherwise must be logical to indicate whether to add group lasso penalties.
intercept	should all intercept be penalized. Disregarded if value is specified.
regression	should all regression parameters be penalized. Disregarded if value is specified.
variance	should all covariance links be penalized. Disregarded if value is specified.
latent	If FALSE, no link related to the latent variable will be penalized. Disregarded if value is specified.
coords	the (spatial) position of the links to penalize

Details

By default categorical variables are penalised using a group lasso penalty. penalize functions can be used to add lasso or/and ridge penalty terms to the model. penaltyNuclear functions can be used to add a nuclear penalty to the model.

Examples

#### lasso penalty ####
set.seed(10)
n <- 500
formula.lvm <- as.formula(paste0("Y~",paste(paste0("X",1:5), collapse = "+")))
lvm.modelSim <- lvm()
regression(lvm.modelSim, formula.lvm) <- as.list( c(rep(0,2),1:3) )
distribution(lvm.modelSim, ~Y) <- normal.lvm(sd = 2)
df.data <- sim(lvm.modelSim,n)

lvm.model <- lvm(formula.lvm)
plvm.model <- penalize(lvm.model)

#### group penalty ####
m <- lvm(Y1 ~ X1+X2+X3+X4)
categorical(m, labels = c("A","B","C")) <- "X1"
categorical(m, labels = c("A","B","C")) <- "X2"

pm <- penalize(m, value = c("Y1~X1","Y1~X3","Y1~X4"))
pm$penalty
pm1 <- penalize(m, value = c("Y1~X1"))
pm1 <- penalize(pm1, value = c("Y1~X2","Y1~X3"))



m <- lvm(list(Y1 ~ X1+X2+X3+X4+eta, Y2 ~ X1+X2+X3+eta, Y3 ~ eta))
categorical(m, labels = c("A","B","C")) <- "X1"
latent(m) <- ~eta

penalize(m)

bozenne/lava.penalty documentation built on May 13, 2019, 1:41 a.m.