R: Model selection

modelSelection

R Documentation

Model selection

Description

Select the best number of breakpoints

Usage

modelSelection(rse, n, c = 2.5, lambdas = NULL, method = c("Birge",
  "Lebarbier"))

Arguments

`rse`	RSE as output by `pruneByDP`
`n`	Length of the profile
`c`	Parameter for the model selection
`lambdas`	A list of candidate values for the calibration of the penalty
`method`	Method to calibrate the constant in the penalty for model selection

Details

This function is not intended to be called directly, but implicitly through jointSeg or PSSeg.

Value

A list with elements

kbest: the best number of breakpoints
lambda: A numerical value, the result of an internal model selection function

Author(s)

Morgane Pierre-Jean and Pierre Neuvial

References

Lebarbier, E. (2005). Detecting multiple change-points in the mean of Gaussian process by model selection. Signal processing, 85(4), 717-736

Birg\'e, L. (2001). Gaussian model selection. J.Eur Math. Soc, 3(3):203-268

Examples


## load known real copy number regions
affyDat <- acnr::loadCnRegionData(dataSet="GSE29172", tumorFraction=1)
sim <- getCopyNumberDataByResampling(1e4, 5, minLength=100, regData=affyDat)
Y <- as.matrix(sim$profile[, "c"])

## Find candidate breakpoints
K <- 50
resRBS <- segmentByRBS(Y, K=K)
## Prune candidate breakpoints
resDP <- pruneByDP(Y, candCP=resRBS$bkp)
selectedModel <- modelSelection(rse=resDP$rse, n=nrow(Y), method="Lebarbier")
str(selectedModel)

## breakpoints of the best model
bestBkp <- resDP$bkp[[selectedModel$kbest]]
print(bestBkp)

## truth
print(sim$bkp)

## Note that all of the above can be done directly using 'PSSeg'
res <- PSSeg(sim$profile, method="RBS", stat="c", K=K)
##  stopifnot(identical(res$bestBkp, bestBkp))