Summary
This version introduces collapse_groups() and friends,
as well as summarize_balances() and
ranked_balances(). It also improves numerical balancing in
fold() which breaks reproducibility.
Changes
Breaking: The numerical balancing
(num_col) in fold() gets multiple
improvements. This breaks reproducibility in some contexts.
Fixes bug with selection of groups to redistribute when
extreme_pairing_levels > 1. The groupings were likely to
be fine, but the fix should give better groupings on average.
When possible, it redistributes the smallest and/or largest group if they are 1 standard deviation from the second smallest/largest group to avoid imbalances due to very small/large scores.
Adds use of extreme triplet grouping when too few grouping
columns are created with extreme pairing. This can lead to an increase
in the number of created fold columns. In some cases, these groupings
may be more balanced than with extreme pairing, but on average extreme
pairing leads to more balanced groupings. See
rearrr::triplet_extremes() for more on extreme triplet
grouping.
Adds argument use_of_triplets in fold()
to allow using extreme triplet grouping instead of extreme pairing or
disabling it completely.
Adds collapse_groups() for collapsing a set of
existing groups into a smaller set of groups. Can balance the new groups
by size and by numeric, categorical and ID columns. The more of these
you balance at a time, the less balanced each will tend to be. Compare
settings by summarizing the balances with
summarize_balances() afterwards. For creating the most
balanced groups, enable auto_tune.
Adds collapse_groups_by_size(),
collapse_groups_by_numeric(),
collapse_groups_by_levels(), and
collapse_groups_by_ids(). These are wrappers of
collapse_groups() for a simplified interface.
Adds summarize_balances() for inspecting the balance
of numeric, categorical, and ID columns in-and-between groups.
Adds ranked_balances() for extracting the
across-group standard deviations of balances from the output of
summarize_balances(). The standard deviations are a measure
of how balanced a split is.
Adds "every" method to grouping functions. Groups
every n data points together.
Prepares package’s tests for
checkmate 2.1.0.
Breaking: Rewrites large parts of the numerical balancing engine
used in fold() and partition(). This produces
different groups in some cases. Outsources extreme pairing to
rearrr::pair_extremes(). Now uses hierarchical shuffling
(rearrr::shuffle_hierarchy()) in partition()
and some cases of fold() (relevant when
extreme_pairing_levels > 1). If you need
reproducibility, the last version prior to this breaking change can be
installed with
devtools::install_github("ludvigolsen/groupdata2@v1.4.2").
Imports rearrr for use in numerical
balancing.
Minor improvements to vignettes.
Adds summarize_group_cols() for finding the number
of groups per fold column along with statistics about the number of rows
per group.
Breaking: Fixes internal sorting of fold columns. This sometimes changes the order of fold columns, compared to the previous version.
Adds tidyr as a required dependency. Previously, it
was suggested.
Breaking: In fold(), the k argument can
now be a multi-element vector with one k (number of folds)
per fold column. This functionality required a minor rewrite, why you
might see interchanged fold column names in comparison to the previous
versions.
Bug fix: In fold() and partition(),
when specifying multiple cat_col columns and
num_col in the same call, it would fail. This now
works.
data.frames (meaning that they are applied group-wise):
fold(), partition(), group(),
group_factor(), splt(),
balance(), upsample(),
downsample(), differs_from_previous(), and
find_missing_starts(). A message is generated once per
session, when the input is grouped, to help users understand why
their code is breaking.checkmate compatibility.
Small speed up of n_dist grouping method.
Adds Zenodo DOI for easier citation.
Adds lifecycle badges to function
documentation.
Adds argument handle_na to
differs_from_previous() and
find_starts().
Bug fix: In grouping functions with method l_starts
and n = "auto", NAs are now replaced by a
unique value before finding group starts. E.g.
c(1,1,1,2,2,NA,NA,4,4) yields 4 groups.
More explicit: the data argument in
fold() and participant takes a data frame, not
a vector.
Possibly breaking change: Adds checkmate input
checks. Improves error messages but also restricts behavior.
Adds xpectr as suggested package. Doubles number of
unit tests.
Adds all_groups_identical() for testing if two
grouping factors contain the same groups, looking only at the group
members, allowing for different group names / identifiers.
Unit tests were made compatible with R versions lower than 3.6.
Adds badges to README, including travis-ci status, AppVeyor status, Codecov, min. required R version, CRAN version and monthly CRAN downloads. Note: Zenodo badge will be added post release.
Bug fix: fold() ungroups dataset before removing
existing fold columns.
Unit tests are skipped on R versions lower than 3.6.
New main function: balance() used for up- and
downsampling of data to balance sample size within categories and IDs.
Thanks for the request from @jjesusfilho (#3).
New wrapper function: upsample() wraps
balance() with size="max".
New wrapper function: downsample() wraps
balance() with size="min".
Adds parameter num_col to fold() and
partition() for balancing on a numeric column.
Adds parameter id_aggregation_fn to
fold() and partition(). Used when balancing on
both id_col and num_col.
Adds helper tool differs_from_previous(). Finds
values in a vector that differs from the previous value by some
threshold. Similar to find_starts().
Adds parameter num_fold_cols to fold().
Useful for creating multiple fold columns for repeated
cross-validation.
Adds parameter unique_fold_cols_only to
fold(). Whether to only include unique fold columns or
not.
Adds parameter max_iters to fold(). How
many times to attempt creating unique fold columns. Note that it is
possible to get fewer fold columns than specified in
num_fold_cols.
Adds parameter parallel to fold(). When
creating multiple unique fold columns, we can run the column
comparisons in parallel. Requires registered parallel backend.
Adds parameter handle_existing_fold_cols to
fold(). When calling fold() on a data frame
that already contains columns with names starting with
".folds", we can either keep them and add more, or replace
them.
Fixed behavior in fold() when k is a percentage
(between 0-1). It is now interpreted as the approximate size of each
fold and used to calculate the number of folds. E.g. k=0.2
will lead to 5 folds.
New main function: partition() - used for creating
balanced partitions by partition sizes.
New method category: l_ methods - n is passed as a
list.
New method: l_sizes - Uses list of group sizes to
create grouping factor. Can be used for partitioning
(e.g. n = c(0.2, 0.3) returns 3 groups with 0.2 (20%), 0.3
(30%) and the exceeding 0.5 (50%) of the data points).
New method: l_starts - Uses list of start positions
to create groups. Define which values from a vector to start a new group
at. Skip to later appearances of a value. Use n = ‘auto’ to
automatically find starts using find_starts().
New helper tool: find_starts() - Finds start
positions in a vector. I.e. values that differ from the previous value.
Get the values or indices of the values. Output can be used as
n in l_starts method.
New helper tool: find_missing_starts() - Returns the
start positions that would be recursively removed when using the
l_starts method with remove_missing_starts set to
TRUE.
Added argument remove_missing_starts to grouping
functions. Recursively remove the starting positions not found with
l_starts method.
New method: primes - similar to
staircase but with primes as steps (e.g. group sizes
2,3,5,7..).
New remainder tool: %primes% - similar to
%staircase% but for the new primes method.
Submitted package to CRAN.
Main functions and tools of this version is
group_factor(), group(), splt(),
fold(), and %staircase%.