plotLambdaVsQF: generates plot of quality function across 'lambda'

In juliancandia/eNetXplorer: Quantitative Exploration of Elastic Net Families for Generalized Linear Models

generates plot of quality function across lambda

Description

Given alpha, this function generates a plot of the quality (objective) function across lambda, which is useful to examine how was the "best lambda" value selected.

Usage

plotLambdaVsQF(x, alpha.index=NULL, xlab="lambda", 
ylab="QF (response vs out-of-bag predicted)", cex.lab=0.95, main=NULL, 
col.main="black", cex.main=0.95, log="x", type="b", ...)

Arguments

x	eNetXplorer object.
alpha.index	Integer indices to select alpha values. Default is 1:length(alpha)
xlab	Custom x-axis label.
ylab	Custom y-axis label.
cex.lab	Axis label size.
main	Custom title.
col.main	Title color.
cex.main	Title size.
log	Log scale axis.
type	Plot type.
...	Additional plotting parameters.

Details

By definition, the "best lambda" value for a given alpha is the one that maximizes the quality function (QF) over the range of lambda values considered. Therefore, QF vs lambda distributions with sharp, narrow, well-defined peaks provide more confidence in the selection of the optimal lambda value than those with less-defined peaks. Sometimes, and particularly for the ridge (alpha=0) solutions, QF is observed to increase or decrease monotonically with lambda over its entire range, causing a boundary lambda value to be selected; we conservatively recommend to disregard alpha-models generated under such circumstances. If interested in investigating further, we suggest to re-run those alpha-models by extending the default range of lambda values (via the argument nlambda.ext) or its density (via the argument nlambda). On occasion, the range of lambda values is effectively limited by convergence issues of the underlying glmnet model; in such scenario, we recommend to increment the maximum allowed number of iterations (via the argument mxit, which is passed on to glmnet.control) or to limit the complexity of the model (e.g. by filtering and reducing the number of features fed into eNetXplorer).

Author(s)

Julian Candia and John S. Tsang
Maintainer: Julian Candia [e-mail]

See Also

eNetXplorer, plot

Examples

data(QuickStartEx)
fit = eNetXplorer(x=QuickStartEx$predictor,y=QuickStartEx$response,
family="gaussian",n_run=20,n_perm_null=10,seed=111)
plot(x=fit,plot.type="lambdaVsQF")
plotLambdaVsQF(x=fit,alpha.index=c(1,3),main="custom title",col.main="red")

juliancandia/eNetXplorer documentation built on May 11, 2023, 1:59 a.m.