impute.learn() now supports training-time storage of
OOD calibration objects through save.ood = TRUE (default),
allowing deployment pipelines to return row-level anomaly summaries in
addition to imputed values.
impute.learn() now accepts optional OOD
weight values at fit time. These can be supplied as named
target weights, are stored in the manifest, and are reused automatically
by impute.ood() when score-time weights are not
supplied.
OOD scoring now returns both a row-level score and a
calibrated score.percentile, with optional target-level
details when return.details = TRUE.
impute.ood() now supports generalized row aggregation
through aggregate and aggregate.args, allowing
users to experiment with row-level anomaly metrics beyond the default
weighted mean. Initial options include weighted mean, weighted
L_p, weighted L_p after a log-tail transform,
top-k, and a bounded product metric on
1 - u_j.
Improvements
OOD scoring is based on masked reconstruction of imputation targets,
using training-time out-of-bag behavior to calibrate target-wise
anomalies.
Numeric targets are scored through reconstruction error, while
factor targets use probability-based surprise when available, with
fallback behavior for cases where probability output is
unavailable.
Unseen factor levels in newdata are now tracked
row-wise and are assigned maximal row-level OOD scores so schema and
category anomalies are easy to identify.
Prediction-time imputation and OOD scoring now share the same
harmonization, iterative sweep, and learner-loading pipeline, including
cache.learners support and richer diagnostics.
When a named OOD weight vector is supplied, weights are
matched by target name; omitted targets receive weight 0,
and extra names are ignored.
Row-level percentile calibration is now rebuilt from saved training
target-wise OOD scores, so score.percentile remains
available for arbitrary target subsets and test-time weight overrides
rather than being limited to the original training-time weighting
scheme.
Row-level percentile calibration is now aligned to the active row
aggregator, so alternate aggregators use their own training-time
reference distribution rather than reusing the weighted-mean
calibration.
If training data are complete and target.mode = "all",
impute.learn() still fits deployable predictive-imputation
learners so later score-time missingness and OOD scoring can be handled
without retraining.
Bug fixes
Fixed cache.learners normalization in the OOD scoring
path.
Fixed row-level OOD calibration reuse after post-fit changes to
saved OOD weights.
Fixed percentile calibration when score-time weights differ from the
original saved weighting scheme.
Fixed handling of partial named weight vectors so omitted targets
are no longer implicitly treated as weight 1.
Improved compatibility behavior for older fitted objects by
preserving raw score output even when legacy objects do not
contain enough saved OOD information to rebuild
score.percentile under the new calibration logic.
Documentation
Expanded the impute.learn help topic to document
impute.ood(), OOD score interpretation, saved weights,
test-time overrides, and unseen-level diagnostics.
Added examples illustrating deployment-oriented use with
target.mode = "all", weighted OOD scoring, and alternate
row aggregators.
randomForestSRC 3.6.1
New features
impute.learn() now supports training-time storage of
out-of-distribution (OOD) calibration references through
save.ood = TRUE (default), enabling a new test-time OOD
scoring workflow via impute.ood() /
impute.ood.rfsrc().
impute.ood() scores new cases by masked reconstruction
across the learned imputation targets and returns row-level OOD scores,
plus score percentiles when the saved row-level calibration is directly
reusable.
OOD scoring now calibrates each target against training-time
out-of-bag reconstruction behavior saved in the imputer manifest,
avoiding a second scoring pass over the training data.
Numeric targets are scored using calibrated absolute reconstruction
error. Factor targets use probability-based surprise when available,
with safe fallback behavior for cases where probability output is
unavailable.
Deployment rows containing previously unseen factor levels are now
flagged explicitly and assigned maximal OOD scores so schema and
category anomalies are easy to identify.
Prediction-time imputation and OOD scoring now share the same
harmonization and learner-loading pipeline, including caching controls
and richer diagnostics for skipped targets, unseen levels, and learner
reloads.
Documentation
Added impute.ood() usage, arguments, return values, and
examples to the existing impute.learn help topic instead of
creating a separate help page.
Documented that target.mode = "all" is the recommended
training configuration when OOD scoring is intended for deployment
use.
Bug fixes
Fixed cache.learners argument normalization in the OOD
scoring path.
randomForestSRC 3.6.0
Changes
Added impute.learn, a predictive imputation framework for test-time
imputation. It learns an imputer from training data that can be saved
and reloaded quickly for compact, low-memory deployment on new
data.
randomForestSRC 3.5.1
Changes
Fixed Uno weights that was breaking for CR.
randomForestSRC 3.5.0
Changes
For survival and competing risk analysis, concordance-based
performance is now computed using Uno inverse-probability-of-censoring
weighting (Uno et al. 2001). This affects all survival performance
values derived from the concordance index, including out-of-bag and test
error rates and variable importance (VIMP). To revert to unweighted
Harrell concordance, set use.uno = FALSE when fitting a survival forest.
Concordance calculations now use an efficient O(n, log n) algorithm
based on a binary indexed tree (Fenwick 1994, Therneau 2024), replacing
the naive O(n^2) pairwise computation for large n. All mclapply() have
been eliminated or replaced by PSOCK providing Windows compatibility as
well.
randomForestSRC 3.4.5
Changes
Change to get.rf.cores() to respect check-as-cran limits.
randomForestSRC 3.4.4
Changes
Improved output for quantile regression. New full sweep option for
imputation via the impute function.
randomForestSRC 3.4.3
Changes
Removed multiplication factor by e^1 for conditional vimp rates for
classification. This was causing vimps greater than one.
randomForestSRC 3.4.2
Changes
Improves “tune.rfsrc” now includes “golden-search”. Improved print
output for grow object. Improved print output for imbalanced forests
“get.imbalanced.performance”.
randomForestSRC 3.4.1
Changes
Bug related to fast/save for classification objects.
randomForestSRC 3.4.0
Changes
Removal of var.select(). It is recommended that users install the
package varPro available on GitHub for variable selection.
randomForestSRC 3.3.1, 3.3.2
Changes
Fix to UBSAN warnings.
randomForestSRC 3.3.1
Changes
Fix to longstanding bug in conditional classification Brier Score.
We were not normalizing conditional values. Addition of case.depth
option.
randomForestSRC 3.3.0
Changes
Resolved an issue with classification where there is a discrepancy
between training and unseen test set class labels. Minor fixes to
subsampling and improved the plot function for subsampling to allow
users to retrieve the plot information for developing custom
figures.
randomForestSRC 3.2.3
Changes
Fix to insecure string warnings on Linux.
randomForestSRC 3.2.2
Changes
Improved the subsampling function. Now users can obtain standard
errors and confidence regions for the generalization error of the
forest. Addition of heavy weighted, and unweighted split rules for mse,
gini, unsupv, and mv cases.
randomForestSRC 3.2.1
Changes
Fix to catestrophic cancellation and buffer overrun in updateCDF(),
when xvar.wt is active with real weights, and mostly when they are very
small values. This bug was introduced in Release 3.1.1.
Fixed long standing bug in mtry when xvar.wt is specified as a real
vector (not uniform, or integers). In such situations, mtry was not
correctly being implemented. Generalized mtry to ytry which is now fully
functional in the multivariate mixed case. Changed suggests from akima
to interp. Updated homepage URL’s.
randomForestSRC 3.1.0
Changes
Added a new way to deal with missing data in prediction where
missing data is assigned randomly by daughter size. Use option
na.action=“na.random”. Added enhanced capability to deal with test data
with factors. Previously test data with factor levels not matching
training data would throw an error. This then required users to apply
hot encoding and other inferior workarounds. These workaround are no
longer needed. Continued improvements to CPU times for training and
prediction. Added a save.memory feature for big survival data to address
slow training times that users were experiencing with large competing
risk and right censored survival data. Added the ability for users to
supply their own sigma matrix for the mahalanobis splitting rule. Added
new features to subsample (used for VIMP confidence intervals). Users
can now specify variables for joint VIMP confidence intervals and
selectively plot confidence intervals using variables of their choosing.
VIMP random now also uses daughter sizes as the secondary threshold
after considering the primary vimp.threshold.
randomForestSRC 3.0.2
Changes
Minor changes to donttest examples that caused issues. Continued
improvements to overall speed of the package.
randomForestSRC 3.0.1
Changes
Minor changes to Rd files. Re-introduction of competing risk
generalized logrank split rule (3.1.1 in the CR paper). This had been
accidentally removed in CRAN build 2_9_0. Fix to DESCRIPTION URL per
Prof. Ripley’s request.
randomForestSRC 3.0.0
Changes
Added the ability to get confidence intervals for out-of-bag error
rates using the subsample function. Significant improvements in CPU
times for restoration of models and prediction with new data. Fixed the
ability to select between in situ ensemble calculation and post-forest
ensemble calculation. This can be useful when doing these calculations
in parallel causes significant contention for memory locations that are
in critical OpenMP regions.
randomForestSRC 2.14.0
Changes
Fix to R_init per Prof. Ripley. Improved output from analysis,
including more reported performance metrics. New get.brier.survival
function for extracting Brier score in survival settings. Importance,
when true, now defaults to anti. Fix to get.tree not playing well with
seed for repeatability of importance.
randomForestSRC 2.13.0
Changes
Fix to missing data segfault related to GitHub issue 112.
Mahalanobis splitting in multivariate regression settings when all
outcomes are real-valued. This can have added benefit when outcomes are
correlated in contrast to the default composite (independent) splitting
rule. Improvements to the speed and output for missForest imputation.
New home page with documentation and vignettes at
https://luminwin.github.io/randomForestSRC
randomForestSRC 2.12.1
Changes
Fix to omp.h declaration before R headers per Prof. Ripley, in
anticipation of clang 13.0.0.
randomForestSRC 2.12.0
Changes
Fixes to variable importance OpenMP threading inefficiencies related
to GitHub issues 95,96,103. Variable importance is now repeatable via
the seed argument, addressing GitHub issues 67,69. Fix to new distance
measure not being normalized, addressing GitHub issue 100. Fixed an
issue related to factors in test data where levels were not matching
traning data. Improvements to user-interface for partial.rfsrc(). Some
improvements in speed and internal logic in the absence of missing
data.
randomForestSRC 2.11.0
Changes
Implementation of rfsrc.anonymous() allowing one to save a model
without the presence of training data, and subsequently predicting
without it, as well. Case specific error and VIMP for classification and
regression.
randomForestSRC 2.10.0
Changes
New additions to the package include sidClustering of unsupervised
data using staggered interaction data (SID) (Mantero and Ishwaran,
2020). Implements the artificial two-class approach of Breiman (2003).
Improved hold out VIMP calculated from the error rate of mini ensembles
of trees (blocks of trees) grown with and without a variable. Applies to
all families. Finally, visualize trees on your Safari or Google Chrome
browser using the new get.tree function for extracting trees from a
forest. Many thanks to @dbarg1 on GitHub for the initial prototype
of this last function.
randomForestSRC 2.9.3
Changes
Bug fixes: Serious impute() issue introduced in the last build (Bug
59 on GitHub). Fix to compile time error on systems that do not support
OpenMP. Here, the OpenMP code that was inadvertently exposed is now
protected when OpenMP support is not detected. Also updated
impute.rfsc.Rd with a salient example.
randomForestSRC 2.9.2
Changes
Bug fixes: Fix to Windows 10 OpenMP stack allocation error
encountered when survival and large n are encountered (Bug 50 on
GitHub). Fix to get.quantile.stat() (Bug 51 on GitHub).
randomForestSRC 2.9.1
Changes
Bug fixes to quantreg and some minor updates to other
functions.
randomForestSRC 2.9.0
Changes
Added new splitrules “auc” and “entropy” for classification.
Improvements to quantile regression. Updates to holdout vimp. Some
function names were changed as a general move towards name uniformity in
the package. Updates to function imbalanced used for class imbalanced
data.
randomForestSRC 2.8.0
Changes
Ensembles in regression now support Greenwald-Khanna approximate
quantile queries via rfsrc(), predict.rfsrc() and the new wrapper
quantileReg.rfsrc(). Related to this, a new split rule “quantile.regr”
has been added. Specifications will be added to the GitHub page,
shortly. Another new wrapper, imbalanced.rfsrc(), implements various
solutions to the two-class imbalanced problem, including the newly
proposed quantile-classifier approach of O’Brien and Ishwaran (2017).
Also includes Breiman’s balanced random forests undersampling of the
majority class. Performance is assesssed using the G-mean, but
misclassification error can be requested. Also, the new parameter
get.tree in predict.rfsrc() allows users to extract the ensembles for a
single tree or subset of trees over the forest. Finally, the default
nodesize for survival and competing risk has been changed to 15.
randomForestSRC 2.7.0
Changes
Three primary additions.
Subsample Forests for VIMP Confidence Intervals: Uses subsampling to
calculate confidence intervals and standard errors for VIMP (variable
importance). Applies to all families.
Tune Random Forest for the optimal mtry and nodesize parameters:
Finds the optimal mtry and nodesize tuning parameter for a random forest
using out-of-bag (OOB) error. Applies to all families.
Fast approximate random forests: Uses subsampling with forest
options set to encourage computational speed. Applies to all
families.
randomForestSRC 2.6.1
Changes
Fix to predict() call not returning ensembles when y-vars not
present. Sorry about that.
randomForestSRC 2.6.0
Changes
Serious improvements to OpenMP performance after addressing issues
related to the blocking of threads during a number of calculations
involving ensembles, importance, forest weights, and partial plots
mostly in big-n data sets.
randomForestSRC 2.5.1
Changes
Addition of configure file to source package allowing more
accessible OpenMP parallel execution on systems that support it.
randomForestSRC 2.5.0
Changes
Introduction of conditional quantiles for a regression forest.
Applies to both univariate and multivariate forests. Can be used for
both training and testing purposes and returns the conditional quantiles
for the target outcomes, and conditional density, which can be used to
calculate conditional moments, such as the mean and standard
deviation.
randomForestSRC 2.4.2
Changes
Bug fixes to partial.rfsrc() on R and C side. Allowance of second
order variable specification in this analysis. Conditional importance
values in classification adjusted by a factor of exp.
Bug fix to unsafe threading in LB-VIMP calculations.
randomForestSRC 2.4.1
Changes
Fix to typedef that breaks Linux. Sorry about that.
randomForestSRC 2.4.0
Changes
Fix to custom splitting family verification and registration
harness. Introduction of bootstrap=“by.user”. Fix to incorrect mapping
of user specified time points to event times when ntime option is used.
It is recommended that the use of this option be avoided. The effect of
discretizing the time values compromises the ensembles. For best
results, all event times must be used. This was and is the default
behaviour. Fix to incorrect passing of time option parameter in
plot.variable. Introduction of partial.rfsrc() to allow direct access to
partial ensembles. Added support for long vectors on native code
side.
randomForestSRC 2.3.0
Changes
Fix to levels.count when ntree=1. Some n-based loop optimization.
Significant improvements in CPU times for restore-predict modes, and
plot.variable(). Consequent changes to forest object, and
incompatibility with objects created with previous versions of the
package. Default is now importance=FALSE in predict.rfsrc().
randomForestSRC 2.2.0
Changes
Bug fix to coerce.factor option via get.xvar.nlevels() and
get.yvar.nlevels() resolved by sending in max instead of number of
levels. Bug fix to VIMP that potentially occurs in OpenMP mode causing
non-zero LB-VIMP. Methodological fix to in-node imputation. and removal
of na.random. Consequent incompatibility with objects created with
previous versions of the package. Fix to rfsrcSyn() bug pertaining to
colnames of test set synthetic features. Introduction of sampsize,
samptype, and case.wt to address imbalanced data sets. Continued
improvements to CPU and memory performance in big-n, big-p, and
big-ntree scenarios.
randomForestSRC 2.1.0
Changes
Change to GROW mode default importance=none and to allow
importance=TRUE. Addition of user trace with time estimates. CPU usage -
code optimization of ensemble calculations. CPU usage - code
optimization of imputation. Fix to R-side parsing of ensembles in
multivariate classification. Change to treat ordered factors under
classification setting instead of regression.
randomForestSRC 2.0.7
Changes
Fix to factor coercion option in responses. Fix to R-side processing
of err.rate and importance in multivariate families with classification.
Update of OPENMP protocols per CRAN recommendation. Expansion of
fast.restore option to omit performance on every tree, and update to
associated Rd file.
randomForestSRC 2.0.5
Changes
Fix to bug in dimensioning of predict object in survival families.
Added documentation for custom splitting.
randomForestSRC 2.0.0
Changes
Multivariate capabilities added. Custom splitting harness
modifications. Redefinition of nodesize to allow terminal nodes less
than said size, subject to the initial test for 2 x nodesize before the
split, maximum depth, and purity. Various bug fixes.
randomForestSRC 1.6.1
Changes
Fix to donttest example in rfsrc.Rd, and other adjustments per CRAN
packaging protocols.
randomForestSRC 1.6.0
Changes
Bug fix to duplicating missingness protocol when restoring a forest.
Added fast.restore option to grow call. Change to pass through xvar.wt
as entered by user. RAM profile reduction in vimp(). Added versioning
checks of forest object, thanks to suggestions by John Ehrlinger. Bug
fix to allow logical responses, treated as reals. User trace
functionality restored.
randomForestSRC 1.5.5
Changes
Bug fix to daughter assignment in classification. Significant RAM
optimization in all modes.
randomForestSRC 1.5.4
Changes
Addition of new function stat.split() for extracting information
from tree node splitting-statistics. Added more functionality to
rfsrcSyn() for fitting synthetic random forests.
randomForestSRC 1.5.3
Changes
Addition of rfsrcSyn() function to grow a synthetic random forest
(RF) using RF machines as synthetic features. Applies only to regression
and classification settings. Used for prediction only.
randomForestSRC 1.5.2
Changes
Fix to non-standard GCC errors and warnings. Fix to bug in split
rules related to omission of missing individuals in the split statistic.
Minor R-side fixes.
randomForestSRC 1.5.1
Changes
Fix to UBSAN warnings. Implementation of new RG protocols.
randomForestSRC 1.5.0
Changes
Significant improvements to CPU and RAM usage profiles in serial and
OpenMP modes of execution. Proximity options allow inbag, OOB and all.
VIMP implements sub-setting and conditional variable importance. NA
options allow the split statistic to be based on non-missing values
only. In addition it allows random assignment of missing values.
randomForestSRC 1.4.0
Changes
Modification of terminal node imputation protocol. We now assign all
individuals the same value rather than sampling from the distribution.
Implementation of split.null option. Implementation of unsupervised
splitting for missing data in impute.rfsrc(). Modification of nimpute
> 1 protocols. In-bag, OOB, and all now depend on the mode. Reduction
in impute memory footprint. Modification of proximity option to allow
in-bag, OOB, and all. Fixed bug in predict involving manual formula
calls. Fixed bug in find.interaction involving specifying covariate
names. Changes relating to Undefined Behaviour Sanitizer.
randomForestSRC 1.3.0
Changes
Initial re-engineering of memory footprint for imputation.
Performance enhancements to split rules.
randomForestSRC 1.2.0
Changes
Competing risks now implements two distinct splitting rules for
identifying short term risks affecting the cause-specific hazard or long
term predictions affecting the cumulative incidence function. The
plot.variable function now returns, and can reuse, a plot.variable data
structure object for user convenience. Thanks to John Ehrlinger for this
improvement. Other minor bug fixes, and enhancements.
randomForestSRC 1.1.0
Changes
OpenMP performance enhancements to ensemble and variable importance
calculations.
randomForestSRC 1.0.2
Changes
Fix to [S] missingness check when all status are non-censored. Fix
to [S] summary imputation of time. Fix to variables used all.trees
output. Fix to manual formula interface. Removed big.data option. Added
ntime option for survival families.
randomForestSRC 1.0.1
Changes
Replaced ‘suggests multicore’ with ‘depends parallel’. Followed
protocol in parallel package for controlling number of cores, via
options(), and environment variables. Reduced [S] memory footprint by
prematurely de-allocating terminal node information.