| rospca | R Documentation |
Robust sparse PCA by the ROSPCA approach of Hubert, Reynkens, Schmitt and
Verdonck (2016): outliers are detected as in robpca(), and sparse
loadings are computed from the clean observations by a grid-search sparse
PCA (Croux, Filzmoser and Fritz, 2013). Sparse loadings are zero for most
variables, which makes the components easier to interpret, for example as
a few emission lines. The computational kernels are written in C++.
rospca(
x,
ncomp = NULL,
lambda = 1,
alpha = 0.75,
ndir = 250,
stand = FALSE,
ngrid = 10,
maxiter = 10,
ncomp_max = 10,
var_explained = 0.8,
k = deprecated()
)
x |
A numeric matrix or data frame, with one observation per row. |
ncomp |
The number of principal components. If |
lambda |
A non-negative number: the sparsity parameter. Default is 1. |
alpha |
The robustness parameter: the fraction of observations the
estimates are based on, between 0.5 and 1. Default is 0.75. The subset
size is |
ndir |
The number of random directions used for the outlyingness.
Default is 250; use |
stand |
A logical value: standardize the variables robustly (median
and |
ngrid |
The number of angles in the grid search. Default is 10. |
maxiter |
The maximum number of grid refinements per component. Default is 10. |
ncomp_max |
The maximum number of components when |
var_explained |
The fraction of variance used to choose |
k |
The algorithm follows the published description in three steps:
Outlier detection. The variables are robustly standardized (median
and Q_n) if stand = TRUE. As in robpca(), the h least
outlying observations form H_0, the number of components
k (ncomp) is chosen (if not given) as the smallest number that
explains var_explained of the variance of H_0, with at most
ncomp_max, and the observations whose orthogonal distance to the
k-dimensional PCA subspace of H_0 is below the cut-off form
H_1 (H_1 = H_0 when ncomp equals the rank of the data, as
the subspace is then the whole space).
Sparsification. The observations of H_1 are standardized and
the sparse loadings are computed by maximizing, component by component,
the variance of the scores minus lambda times the L_1 norm of
the loadings. Variables with zero loadings on all components are set
aside; the observations whose orthogonal distance to the sparse
subspace is below the cut-off form H_2 (H_2 = H_1 when the
sparse subspace contains all the observations), and the sparse loadings
are recomputed from H_2, standardized in turn.
Eigenvalues and center. The eigenvalues are first estimated by the
squared Q_n of the scores of H_2. The h observations of
H_2 with the smallest score distances form H_3; the center
is their mean,
and the final eigenvalues are the variances of their scores. The
components are sorted by decreasing eigenvalue.
Larger lambda values give sparser loadings. The loadings apply to the
data standardized by center and scale; they are unit vectors, but
those of different components are not exactly orthogonal. The eigenvalues
are the variances of the scores in these standardized units. lambda is best chosen
by validation, for example with step_rospca() and tune.
This is an independent implementation of the published description, not
a port of the rospca code, so results can differ from those of
rospca::rospca().
An object of class specproc_rospca (inheriting from
specproc_robpca), with the components described in robpca(), and:
scale: the scales that standardize the variables before projection:
the robust scale (Q_n, if stand = TRUE) times the standard
deviation in H_2.
lambda: the sparsity parameter.
H2, H3: logical vectors of the observations in these subsets.
Christian L. Goueguel
Hubert, M., Reynkens, T., Schmitt, E., Verdonck, T. (2016). Sparse PCA for high-dimensional data with outliers. Technometrics, 58(4):424-434.
Croux, C., Filzmoser, P., Fritz, H. (2013). Robust sparse principal component analysis. Technometrics, 55(2):202-214.
robpca(), step_rospca(), plot_outlier_map()
set.seed(1)
# the 380-430 nm window (Ca II H and K lines)
wl <- suppressWarnings(as.numeric(names(forageLIBS)))
forageLIBS[, which(wl > 380 & wl < 430)] |>
center() |>
rospca() |>
print()
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.