| NEWS | R Documentation |
Fix an issue with printing of tied p-values.
pruning of modelparty objects failed to get the fitted slot right.
In R-devel, c(<factor>) now returns factors, rendering code in .simplify_pred overly pedantic.
NAMESPACE fixes: party is only suggested.
Remove warning about response not being a factor in
predict.cforest. Reported by Stephen Milborrow.
Trying to split in a variable where all observations were missing nevertheless produced a split, as reported by Kevin Ummel.
update reference output, fix RNGversion
varimp runs in parallel mode, optionally.
weights in cforest can now be used
to specify a matrix of weights (number of observations
times number of trees) to be used for tree induction (this was
always possible in party::cforest. This was advertised in
the documentation but actually not implemented so far.
predict did not pay attention to xlev; this caused
problems when empty factor levels were removed prior to tree fitting.
nodeprune may have got fitted terminal node numbers wrong,
spotted by Jason Parker.
In mob() using the cluster argument with a factor
variable sometimes lead to NAs in the covariance matrix estimate
if empty categories occured in subgroups. The problem had been introduced
in version 1.2-0 and has been fixed now.
Methods for the sctest generic from the strucchange
package are now dynamically registered if strucchange is attached.
Alternatively, the methods can be called directly using their full names
sctest.constparty and sctest.modelparty.
The prune.modelparty function is now fully exported but
it is also registered with the prune generic from rpart.
New scale argument for predict in cforest.
For simple regression forests, predicting
the conditional mean by nearest neighbor weights with
scale = TRUE is now
equivalent to the aggregation of means. The unscaled version
proposed in <doi:10.1002/sim.1593> can be obtained with
scale = FALSE.
Bug fix for case weights in mob() in previous version (1.2-0)
introduced a bug in the handling of proportionality weights. Both cases
are handled correctly now.
glmtree can now handle caseweights = TRUE correctly
for vcov other than the default "opg". Internally,
the glm objects are adjusted by correcting the dispersion
estimate and the degrees of freedom.
lookahead did not work in the presence of missing values.
Calling partykit::ctree did not work when partykit was not
attached.
node_inner now allows to set a different gpar(fontsize = ...)
in the inner nodes compared to the overall tree.
splittest asked for Monte-Carlo p-values, even
when the test statistic was used as criterion.
We welcome Heidi Seibold as co-author!
Internal re-organisation for ctree by means of new extensible
tree infrastructure (available in extree_data and
extree_fit). Certain parts of the new infrastructure
are still experimental. ctree is fully backward
compatible.
Use libcoin for computing linear test statistics
and p-values for ctree.
Use inum for binning (the new nmax argument).
Quadratic test statistics for splitpoint selection are now available for ctree via
ctree_control(splitstat = "quadratic").
Maximally selected test statistics for variable selection are now available for ctree via
ctree_control(splittest = TRUE).
Missing values can be treated as a separate category, also
for splits in numeric variables in ctree via
ctree_control(MIA = TRUE).
Permutation variable importance, including conditional variable importance, was added to partykit.
New offset argument in ctree.
New get_paths for computing paths to nodes.
node_barplot gained a text argument
that can be used to draw text labels for the
percentages displayed.
The margins used in plot.party can now
also be set by the user.
Bug fix in mob() if weights are used and caseweights = TRUE (the default).
The statistics for the parameter instability tests were computed incorrectly and
consequently the selection of splitting variables and also the stopping criterion
were affected/incorrect.
Avoid log(p) values of -Inf inside mob() by replacing weighted averaging
with naive averaging in the response surface regression output in case
the p values are below machine precision.
The as.party method for rpart objects
without any splits only returned a naked partynode
rather than a full party. This has been corrected
now.
nodeapply did not produce the same results for permutations of ids.
Spotted by Heidi Seibold.
Out-of-bag predictions in predict.cforest were incorrect.
perm in predict was only considered when
newdata was given. Spotted by Heidi Seibold.
Don't try to search for binary splits in unordered factors with more than 31
levels. This potentially caused an integer overrun in previous versions.
party::ctree() uses an approximation for binary split
searches in unordered factors; thus, using party might be an
alternative.
Proper support of quasi-families in glmtree
and hence palmtree.
NA handling by following the majority was
potentially incorrect in ctree.
Minor speed improvements.
Breaking ties before variable selection was suboptimal for very small log-p-values.
Added a new function palmtree that fits partially
additive (generalized) linear model trees. These employ
model-based recursive partitioning (mob) based on (generalized)
linear models with some local (i.e., leaf-specific) and
some global (i.e., constant throughout the tree) regression
coefficients.
Splits in ordinal variables are now represented correctly
in the (still internal) list.rule method.
Kaplan-Meier curves in "constparty" trees were
plotted incorrectly due to use a wrong scaling of the
x-axis. Spotted by Peter Calhoun <calhoun.peter@gmail.com>.
Use quote(stats::model.frame) instead of
as.name("model.frame").
The as.party methods for rpart and
Weka_tree now have a data = TRUE argument
so that by default the data is preserved in the party
object (instead of an empty model frame).
The predict method for cforest objects
did not work for one-row data frames, fixed now.
Added rot and just arguments to node_barplot
for more fine control of x-axis labeling (e.g., with 45 degree
rotation).
The partykit package has now been published in Journal of Machine Learning Research, 16, 3905-3909. http://jmlr.org/papers/v16/hothorn15a.html
Added support for setting background in panel functions.
The as.list() method for partynode objects
erroneously created an object thisnode in the calling
environment which is avoided now.
Bug fix in plot() method for constparty objects.
In the previous partykit version clipping was accidentally
also applied to the axes labels.
For constparty objects plot(..., type = "simple")
did not work correctly whereas plot(as.simpleparty(...))
yielded the desired visualization. Now internally as.simpleparty()
is called also in the former case.
The as.simpleparty() method now preserves p-values from
constparty objects (if any).
Added a getCall() method for "party" objects.
In the predict() method for "lmtree" and "glmtree"
objects the offset (if any) was sometimes ignored. It is
now always used in the prediction.
Import logrank_trafo from coin.
nodeprune(..., ids = 1) did not prune the tree to the root
node. Fixed now.
predict.cforest used na.omit instead of na.pass.
predict.party now features new perm argument for
permuting splits in specific variables (useful for computing
permutation variable importances).
NAMESPACE updates.
The support for (generalized) linear model trees with just
a constant regressor has been improved. Now lmtree(y ~ x1 + x2)
is short for lmtree(y ~ 1 | x1 + x2), analogously for glmtree().
Plotting now also works properly in this case.
The as.party() method for "rpart" objects did not
work if one of the partitioning variables was a "character"
variable rather than a "factor". A suitable work-around has
been added.
The node_barplot() panel function can now also be used
for multivariate responses, e.g., when all responses are numeric and
on the same scale.
The package now also includes a new data set HuntingSpiders
which is essentially a copy of the spider data from the package
mvpart that is currently archived on CRAN. The documentation has
been improved somewhat and is likely to change further to explain how
the data has been transformed in De'ath (2002).
The survival tree example for the GBSG2 data was broken due to the response being (incorrectly) also part of the explanatory variables. Fixed by using the latest Formula package (at least version 1.2-1).
Version 1.0-0 published. This version is described in the MLOSS paper accepted for publication by the Journal of Machine Learning Research today.
The unbiased version of cforest (with replace = FALSE) is
now the default (as in party).
Register all S3 methods in NAMESPACE.
Extended mob() interface by a cluster argument. This can be
a vector (numeric, integer, factor) with cluster IDs that are
then passed on to the 'fit' function (if supported) and used
for clustering the covariance matrix in the parameter stability
tests. lmtree() and glmtree() hence both gained a cluster
argument which is used only for cluster covariances but not
for the model estimation (i.e., corresponding to a working
independence model).
Optionally, the parameters' variance-covariance matrix in mob()
can now be estimated by the sandwich matrix instead of the default
outer-product-of-gradients (OPG) matrix or the information matrix.
Reimplementation of cforest() available with extended
prediction facilities. Both the internal representation and the user interface
are still under development are likely to change in future versions.
Added multicore support to mob(), ctree(), and cforest().
If control argument cores is specified (e.g., cores = 4) then the
search for the best variable or split point (often involving numerous model fits in
mob() or resampling in ctree()) is carried out using parallel::mclapply()
rathern than sequential for() or sapply(). Additionally, other
applyfuns can be provided, e.g., using networks of workstations etc.
Bug fix in mob() that occurred when regressor variables and
partitioning variables overlapped and were not sorted in the
underlying model frame.
mvpart was archived 2014-12-15.
Fixed an uninitialized memory issue reported by valgrind.
partykit now depends on R version >= 3.1.0 in order to import the
depth() generic from the grid package.
The print methods for party/partynode objects with only a root node
was modified. Now, the terminal panel function is also applied
if there is only a root node (while previously it was not).
ctree() now catches sum(weights) <= 1 situations before they
lead to an error.
Code from suggested packages is included by using :: syntax as
required by recent R versions.
Argument ytrafo of ctree() can now be a function which will be
updated in every node.
A small demo briefly illustrating some memory and speed properties
has been added. It can be run interactively via
demo("memory-speed", package = "partykit").
Section 3 of the "constparty" vignette now shows how properties of a new tree algorithm can be assessed by using partykit building blocks.
Major improved version of partykit. The previously existing functions in the package were tested and enhanced, new functions and extensive vignettes added.
Extended and improved introductory documentation. The basic classes
and class constructors partynode/partysplit/party are introduced in
much more detail now in vignette("partykit", package = "partykit").
The class constparty (inheriting from party) for representing party
objects with constant fits in the nodes (along with coercion methods
for rpart, J48, etc.) is now described in more detail in the new
vignette("constparty", package = "partykit").
The package now includes a reimplementation of the model-based
recursive partitioning algorithm (MOB) using partykit infrastructure.
The generic algorithm, the corresponding convenience interfaces
lmtree() and glmtree() as well as various illustrations and possible
extensions are described in detail in the new
vignette("mob", package = "partykit").
Improved implementation of conditional inference trees (CTree), see
the new vignette("ctree", package = "partykit") for details.
New nodeprune() generic for pruning nodes in all party trees and
partynode objects.
Deal with empty levels in ctree() for teststat = "quad"
(bug reported by Wei-Yin Loh <loh_at_stat.wisc.edu>).
In predict() method for constparty objects, type = "prob" now returns
ECDF for numeric responses and type = "response" the (weighted) mean.
New panel function node_ecdf() for plotting empirical cumulative
distribution functions in the terminal leaves of constparty trees.
Bug fix in as.party() method for J48 trees with ordered factors.
Fix C code problems reported by clang under OS X.
Added node_surv() for plotting survival ctrees. Accompanying
infrastructure for survival trees was enhanced.
ctree() now checks for (and does not allow) x >= max(x) splits.
Added ipred to the list of suggested packages due to usage of GlaucomaM and GBSG2 data in tests/examples.
The node_terminal() panel-generating function is now customizable
by a FUN argument that is passed to formatinfo().
The plot() method for simpleparty object now sets up a formatting
function passed to formatinfo(), both in print() and plot().
Fixed bug in pmmlTreeModel() for processing label IDS in splits when
not all levels are present.
Cleaned up unused variables in C code and partial argument matching in R code.
First CRAN release.
See vignette("partykit", package = "partykit") for a (somewhat rough)
introduction to the package and its classes/methods.