splithalf() sampling for
stability(). In case of an uneven number of observations in
the learning sample some small overlap in the two splitted halves would
occur. This is avoided now by dropping one random observation now and
restricting both halves to be always of the same size. Reported by
Constantin Wiegand along with two further small improvements: Warnings
about lack of variance in predictions on learning samples and avoiding
manual triggering of the garbage collector.The as.stabletree() method for
RandomForest objects (party package) is now
registered as an S3 (rather than S4) method.
Environment .stabEnv is only used internally by
stabletree() and hence not exported anymore in
NAMESPACE.
Z takes over maintenance from MP.
Fix checks for sampler specifications where a
logical of length 2 is now correctly aggregated with
all().
Changed default sampling method in stabletree() from
bootstrap() to subsampling() with default
fraction of v = 0.632.
as.stabletree() coercion generic added which allows
to coerce a randomForest (randomForest package),
RandomForest (party package), cforest
(partykit package) or ranger (ranger
package) to a stabletree object.
Added a vignette on the variable and cutpoint selection analysis of random forests.
Project stablelearner has been launched and a stable
version of the package has been uploaded to CRAN.
stability() is available to estimate the stability
of the results from a given supervised statistical learning
method.
stabletree() is available to estimate the stability
of the results from recursive partitioning.