The hardware and bandwidth for this mirror is donated by dogado GmbH, the Webhosting and Full Service-Cloud Provider. Check out our Wordpress Tutorial.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]dogado.de.
This release is a redesign of the package. Code written for 0.2.x
must be updated; see
vignette("migration", package = "MFF").
mff(), tune.mff(),
model.train(), boot.train(), the
mff class and its print(),
summary(), plot() and predict()
methods.foreach and doParallel are no longer
dependencies; parallel tuning and bootstrap generation use socket
clusters from base R’s parallel package.mff_pool() creates an mff_prediction_pool
from aligned validation and test prediction matrices produced by any
software; model_train() and boot_train()
return the same class.fit_mff() fits one configuration (class
mff_fit); tune_mff() searches a grid and
returns c("mff_tuned", "mff_fit").selected_configuration() returns only the parameters
that apply to the method and prints them compactly; pool summaries print
as a short table.validation_predictions(),
validation_observed(), test_predictions(),
test_observed(), candidate_names(),
candidate_info(), pool_info(),
validation_scores(), selected_function(),
selected_configuration(), search_results() and
fit_diagnostics().pool_info() returns a list of class
mff_pool_info. Its print method shows the task, pool type,
indexes, split sizes, resampling and the learner settings in compact
form; the full components remain available through $ and
[[.print(), summary(),
predict() (returns numeric predictions),
weights(), coef() and plot()
(weights, heatmap, validation and test scores, observed versus
predicted, series, search). Score plots mark the function selected on
validation data; the search plot labels every configuration, marks the
selected one and lists failed configurations; heatmap labels and cell
text adapt to names and colors.mff_split() defines explicit training, validation and
test rows; model matrices, factor levels and offsets are learned from
training rows only.model_train() accepts a selected set of learners with
per-learner controls.mff_learner() wraps any regression method (a
fit(x, y, ...) and a predict(object, newx)
function) for model_train() and boot_train().
model_train() accepts named lists of learners, so one
algorithm can enter several times with different controls, and
boot_train() accepts any built-in or user learner instead
of only lm.model_train(resampling = mff_cv()) builds a
cross-validated pool: training rows are predicted out of fold and used
as validation predictions, and test predictions come from candidates
refitted on all training rows. Encoding and learner tuning are repeated
within each fold. mff_cv(fold_id = ) accepts user-defined
folds, and mff_split() accepts
validation = NULL.as_mff_pool() converts resampling results of other
frameworks: named lists of caret::train() models and
caretEnsemble caretList objects (heterogeneous pools),
tidymodels tune_grid() results (hyperparameter pools, with
test predictions from refitted workflows) and stacks data stacks
(validation-only pools). The frameworks are optional.rolling_train() builds forecasting pools by
rolling-origin evaluation: each target is forecast horizon
steps ahead from models fitted only to the observations up to its origin
(expanding or sliding windows). The validation block ends at the origin
of the first test target, so for horizon > 1 the
horizon - 1 target times between the blocks are skipped and
no validation outcome is observed after a test origin. Models are
functions function(y, h) and may return forecast package
objects. as_mff_pool() converts fable forecasts of a single
series from stretched tsibbles with the same validation boundary.nstart is honored by every clustering method, selecting
the start with the lowest clustering objective; reltol
controls convergence tolerance.fit_diagnostics()$convergence reports
"converged", "iteration_limit" or
"backend_warning". A start counts as converged only when it
meets the backend stopping rule before iter_max without
warnings; kmeans status is read from its convergence code.ppclust::pfcm() (version 1.1.0.1) is no longer used:
it raises the sum in equation (21) to the power 1/(m - 1) instead of
each term, so memberships are wrong and do not sum to one for m other
than 2, and it assigns zero-distance points to the last cluster. ppclust
is no longer a dependency.fit_mff(), tune_mff() and the internal
PFCM routine require eta > 1, since the typicality
update of Pal et al. (2005) contains the exponent 1/(eta - 1). Earlier
checks accepted any positive eta: values equal to 1 gave
crisp 0/1 typicalities, and values below 1 gave typicalities that
increase with distance, both without an error. The default
eta = 2 is unchanged.abs(actual) + 1e-8,
so that the denominator cannot vanish for negative actual values.
fit_mff() stops with a message when the selection metric is
not finite for any meta fuzzy function, and tune_mff()
reports this message for failed grids.fit_mff() requires fewer clusters
than candidates and stops with a message otherwise;
tune_mff() records such configurations as failed.rolling_train() checks
n_validation >= 2 and a strictly increasing
index before any model is fitted, and
model_train() and boot_train() reject
duplicated learner control names before training.reltol.beta and covariance weighting
gamma as new arguments of fit_mff() and
tune_mff(); the default GK tolerance is 1e-3 on the largest
membership change. Memberships are normalized on the log scale, which
avoids overflow for small m, and a non-finite update or a
numerically zero covariance matrix fails the start explicitly.
ppclust::gk() (version 1.1.0.1) is no longer used: its
distance is (2*pi)^(N/2) * det(F)^(1/2) times the
Mahalanobis term divided by the cluster proportion, with N the number of
objects, rather than the GK distance det(F)^(1/n) times the
Mahalanobis term, with n the dimension; (2*pi)^(N/2)
overflows for more than 772 objects; and nonfinite memberships made it
return the previous memberships as if converged.mff_diamonds: validation and test
predictions of seven regression learners for diamond price on a random
split of 2000 rows of ggplot2::diamonds.
mff_example() now builds this seven-candidate pool; the
previous example split trees by row order, so validation
and test rows extrapolated and MFF effectively selected one
candidate.evaluate().tune.mff() grid searches by making all
internal functions required by worker processes available in their
execution environments.mff() now requires at least two clusters/functions,
consistently with the search space used by tune.mff().mff() cluster-count boundary
and equality of sequential and parallel k-means tuning results under a
common seed.plot.mff(type = "weight_heatmap") for a labelled
heatmap of the complete candidate-model by meta-function weight
matrix.predict.mff() now verifies candidate-model identities
and column order.evaluate() now rejects mismatched, empty, malformed,
and non-finite inputs.testthat suite, and
print(), summary() and plot()
methods for fitted and tuned MFF objects.model.train() to
one thread.These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.
Health stats visible at Monitor.