optLambdas {multiridge}

Find optimal ridge penalties.

Description

Optimizes a cross-validated score w.r.t. ridge penalties for multiple data blocks.

Usage

optLambdas(penaltiesinit = NULL, XXblocks, Y, X1 = NULL, pairing = NULL, folds,
  intercept = ifelse(is(Y, "Surv"), FALSE, TRUE), frac1 = NULL, score = "loglik",
  model = NULL, epsIWLS = 0.001, maxItrIWLS = 25, traceCV = TRUE, reltol = 1e-04,
  optmethod = ifelse(length(penaltiesinit) == 1, "Brent", "Nelder-Mead"), maxItropt = 500,
  save = FALSE, parallel = FALSE, fixedpen = NULL, fixedseed = TRUE)

Arguments

penaltiesinit

Numeric vector. Initial values for penaltyparameters. May be obtained from fastCV2.

XXblocks

List of nxn matrices. Usually output of createXXblocks.

Y

Response vector: numeric, binary, factor or survival.

X1

Matrix. Dimension n x p_0, p_0 < n, representing unpenalized covariates

pairing

Numerical vector of length 3 or NULL when pairs are absent. Represents the indices (in XXblocks) of the two data blocks involved in pairing, plus the index of the paired block.

folds

List, containing the splits of the samples. Usually obtained by CVfolds

intercept

Boolean. Should an intercept be included?

frac1

Scalar. Prior fraction of cases. Only relevant for model=" logistic".

score

Character. See Details.

model

Character. Any of c("linear", "logistic", "cox"). Is inferred from Y when NULL.

epsIWLS

Scalar. Numerical bound for IWLS convergence.

maxItrIWLS

Integer. Maximum number of iterations used in IWLS.

traceCV

Boolean. Should the output of the IWLS algorithm be traced?

reltol

Scalar. Relative tolerance for optimization methods.

optmethod

Character. Optimization method. Any of the methods c("Brent", "Nelder-Mead", "Sann") may be used, but "Nelder-Mead" is generally recommended. Other unconstrained methods offered by optim may also be used, but have not been tested.

maxItropt

Integer. Maximum number of iterations for optmethod.

save

Boolean. If TRUE appends the penalties and resulting CVscore to global variable allscores

parallel

Boolean. Should computation be done in parallel? If TRUE, requires to run setupParallel first.

fixedpen

Integer vector or NULL. Contains indices of data types of which penalty is fixed to the corresponding value in penaltiesinit.

fixedseed

Boolean. Should the initialization be fixed? For reproducibility.

Details

See Scoring for details on score. We highly recommend to use smooth scoring functions, in particular "loglik". For ranking-based criteria like auc and cindex we advise to use repeated CV (see CVfolds) to avoid ending up in any of the many local optima.

Value

List, with components:

optres

Output of the optimizer

optpen

Vector with determined optimal penalties

allsc

Matrix with CV scores for all penalty parameter configurations used by the optimizer

See Also

optLambdasWrap for i) (recommended) optimization in two steps: first global, then local; and ii) sequential optimization when some data types are preferred over others. fastCV2 for initialization of penalties. A full demo and data are available from:
https://drive.google.com/open?id=1NUfeOtN8-KZ8A2HZzveG506nBwgW64e4

Examples

data(dataXXmirmeth)
resp <- dataXXmirmeth[[1]]
XXmirmeth <- dataXXmirmeth[[2]]

# Find initial lambdas: fast CV per data block separately.
cvperblock2 <- fastCV2(XXblocks=XXmirmeth,Y=resp,kfold=10,fixedfolds = TRUE)
lambdas <- cvperblock2$lambdas

# Create (repeated) CV-splits of the data.
leftout <- CVfolds(Y=resp,kfold=10,nrepeat=3,fixedfolds = TRUE)

# One-pass optimization
# Increase the number of iterations for optimal results
jointlambdas <- optLambdas(penaltiesinit=lambdas, XXblocks=XXmirmeth,Y=resp,
folds=leftout,score="loglik",save=T,maxItropt=5)