C5.0() and predict() now pass the data
to the C code as R vectors. Before, the data were written as the text
files of the command-line program (.names,
.data, .cases, .costs) and parsed
again by the C code. For data that the text files could hold, the models
and predictions are the same as before. The text interface had these
problems, which are now fixed:
"?" became a missing value, a level "N/A"
became “not applicable”, and whitespace in names and levels was trimmed
or collapsed, so levels such as "a b" and
"a b" were merged.case weight was used as case weights
(and not as a predictor), even when no weights were
given."c,d") were split
into several levels, and the fit failed.continuous,
ignore, label, or date was read
as a type.attributes included or
attributes excluded were read as instructions, and names
and values longer than 1000 characters gave an error."NA" as a level, and their
levels were in the order of appearance. They now use sorted levels (in a
locale-independent order). Models with character predictors can
therefore differ slightly from before.Off-diagonal values of a cost matrix that are not greater than 1
are no longer dropped. Before, C5.0() passed only costs
greater than 1 to the C code, so a cost of 0.5 (for example) had no
effect. Models fitted with such costs will differ. Also, the rows of the
cost matrix are now matched to the class levels by their row names;
before, the row names were ignored and the rows were assumed to be in
the order of the column names. See the new “Costs” section of
?C5.0 for how costs are used.
Cost matrices are now checked before the fit: the values must be
finite and non-negative, and with two classes the off-diagonal values
must be greater than 0. (A zero cost with two classes makes the C code
build a broken model that crashes in predict().)
predict() now treats factor or character values in
newdata that were not in the training data as missing
values, with a warning. Before, it stopped with an error (“bad value of
… Error limit exceeded”) when the model was fitted with x
and y.
C5.0() now gives clear errors for logical or date
predictors, infinite values in the predictors or weights, a
predictor with the same name as the outcome label
(C5.0Control(label = )), duplicated column names, an
outcome with no levels, and y or weights with
the wrong length.
Fitted objects have two new elements: spec (the
predictor names, types, and levels) and format_version. The
names element (the text of the .names file) is
still there for code that reads it, but it no longer has comment lines
with the R version and a time stamp. Fitted objects from earlier
versions can still be used with predict(); for them, the
new data are still written as text with
Cubist::makeDataFile(), so changes to that function in
Cubist also affect predictions from those old objects.
An outcome with a single level now gives a one-leaf model that
predicts that level (with probability 1). In version 0.2.0 the C code
failed but C5.0() still returned an object.
The internal helper that as.party() uses to get the
training data is no longer registered as a model.frame()
method for C5.0 fits. model.frame() again uses the default
method and returns a standard model frame, as it did in version 0.2.0.
This fixes model.response(model.frame(fit)), which broke
ModTools.
For new fits with case weights, C5imp() no longer
lists case weight as a predictor.
C5.0() no longer limits trials to 100.
Any whole number >= 1 is allowed; the C code already allocated its
per-trial storage at run time, so the cap existed only in R. Models with
100 or fewer trials are unchanged. Note that some downstream packages
have their own cap of 100: parsnip (boost_tree() C5.0
engine tuning range and C5_rules()), rules
(C5_rules fit wrapper and tuning range), and caret (random
search over 1:100).
print.C5.0() now says “no early stopping for
boosting” when earlyStopping = FALSE. Before, it said
“early stopping for boosting”, which read as if early stopping were
turned on.
Added strip_time_stamps option to
C5.0Control() (default TRUE) to remove
timestamps from model output, making results reproducible across
runs.
Switched from base R stop() and
warning() to cli’s cli_abort() and
cli_warn() for improved error messages with better
formatting and context.
Fixed compiler warnings for unused variables in C source code.
Maintenance release to fix CRAN issues related to compiler issues.
Updated GHA and tests
Maintenance release to fix CRAN issues by adding
rmarkdown to Suggests.
Fixed CRAN issues related to undefined behavior in the C code
when requesting rules with C5.0(rules = TRUE).
Fixed CRAN issues for GCC 10 -fno-common
flag.
Added a NEWS.md file to track changes to the
package.
Removed churn data in favor of the version in the
modeldata package.