View source: R/robust_rmsecv.R
| robust_rmsecv | R Documentation |
Robust root mean squared error of cross-validation (R-RMSECV) of the
robust PLS (rsimpls()) or robust PCR (rpcr()) models with 1 to
ncomp_max components, and the robust component selection (RCS) criterion of
Engelen and Hubert (2005), to choose the number of components when the
calibration data contain outliers.
robust_rmsecv(
x,
y,
method = c("rsimpls", "rpcr"),
ncomp_max = 10,
folds = 10,
gamma = 0.5,
alpha = 0.75,
ndir = 250,
nsamp = 500,
...
)
x |
A numeric matrix or data frame of the predictors (spectra), one observation per row. |
y |
A numeric vector, matrix or data frame of the responses, with one row per observation. |
method |
The model: |
ncomp_max |
The largest number of components. Default is 10. |
folds |
The number of folds of the cross-validation, from 2 to the number of observations (leave-one-out). Default is 10. |
gamma |
The weight of the cross-validated error in the RCS criterion, between 0 and 1. Default is 0.5. |
alpha |
The robustness parameter of |
ndir |
The number of random directions of the outlyingness in
|
nsamp |
The number of random subsets of FAST-MCD (and, for
|
... |
Not used. |
The squared prediction errors of outliers would dominate the usual
RMSECV. The robust RMSECV leaves them out: the model is first fitted to
all the observations, and only those regular in every model with 1 to
ncomp_max components (residual distance below its cut-off) enter the error,
R\text{-}RMSECV_k = \sqrt{\frac{1}{q \sum_i w_i} \sum_i w_i
\|y_i - \hat y_{-i,k}\|^2},
where w_i is 1 for these observations and 0 for the others, and
\hat y_{-i,k} is the prediction of observation i by the model
with k components fitted without it.
The robust root mean squared error of the fit to the calibration data,
RMSE (the square root of the robust residual sum of squares, R-RSS),
decreases with the number of components, while the cross-validated error
eventually increases. The RCS criterion combines them,
RCS_k = \sqrt{\gamma \, R\text{-}RMSECV_k^2 + (1 - \gamma) \,
R\text{-}RSS_k},
and its first local minimum, or the point after which it decreases
little, suggests the number of components. gamma = 1 gives the robust
RMSECV, and gamma = 0 the R-RSS.
Differences from the paper. The paper computes the leave-one-out
errors with a fast approximation that updates the robust fits instead of
refitting them. Here, each fold is refitted. Leave-one-out
(folds = nrow(x)) is then slow, and K-fold cross-validation (default
10 folds) is used instead.
The folds are fitted in parallel when a parallel plan is set with
future::plan() (and the future.apply package is installed). The folds
and the fits are random: use set.seed() for reproducible results.
A tibble with one row per number of components: ncomp, the
robust R2 and RMSE of the fit to all the observations (as the
components of rsimpls() and rpcr()), the robust RMSECV and the
RCS criterion. The weights w_i are its "weights" attribute.
Christian L. Goueguel
Engelen, S., Hubert, M. (2005). Fast model selection for robust calibration methods. Analytica Chimica Acta, 544(1-2):219-228.
rsimpls(), rpcr()
data(forageLIBS)
wl <- suppressWarnings(as.numeric(names(forageLIBS)))
spectra <- forageLIBS[which(wl > 380 & wl < 430)]
set.seed(1)
robust_rmsecv(spectra[1:300, ], forageLIBS$Ca[1:300], ncomp_max = 8, folds = 5)
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.