Machine Learning in R
Adjust behavior of "positive" arg for classif.logreg (#2846)
Consistent naming for dummy feature encoding of variables with different levels count (#2847)
Remove {nodeHarvest} learners (#2841)
Remove {rknn} learner (#2842)
Remove all {DiscriMiner} learners (#2840)
Remove {extraTrees} learner (#2839)
Remove depcrecated {rrlda} learner
Resolve some {ggplot2} deprecation warnings
Fixed information.gain filter calculation.
Before, chi.squared was calculated even though information.gain was requested due to a glitch in the filter naming (#2816, @jokokojote)
Make helpLearnerParam()'s HTML parsing more robust (#2843)
Add HTML5 support for help pages
FSelectoRcpp::relief(). This C++ based implementation of the RelieF filter algorithm is way faster than the Java based one from the {FSelector} package (#2804)FilterWrapper objectsfix.factors.prediction = TRUE causes the generation of NAs for new factor levels in prediction (@jakob-r, #2794)newdata (@jakob-r, #2794)praznik_MRMR: Remove handling of survival tasks (#2790, @bommert)objective default from reg:linear (deprecated) to reg:squarederror
blocking was set in the Task but blocking.cv was not set within `makeResampleDesc() (#2788)generateLearningCurveData() (#2768)getFeatureImportance(): Account for feature importance weight of linear xgboost modelss did not match the learner note) (#2747)createSpatialResamplingPlots(). The package caused issues on R-devel. In addition users should set custom themes by themselves.getNestedTuneResultsOptPathDf() (#2754)regr_slim learner due to pkg (flare) being orphaned on CRANclValid::dunn and its tests (package orphaned) (#2742)tuneThreshold() now accounts for the direction of the measure.
Beforehand, the performance measure was always minimized (#2732).more.args for simple filters (@annette987, #2709)print.FeatSelResult() when bits.to.features is used in selectFeatures() (#2721)getFeatureImportance() (#2708)pkgdown: Move changelog to Appendix
Account for {checkmate} v2.0.0 update (#2734)
Refactor function calls from packages (<pkg::fun>) within ParamSets (#2730) to avoid errors in listLearners() if those pkgs are not installed
listLearners() should not fail if a package is not installed (#2717)
n.show argument had no effect in plotFilterValues(). Thanks @albersonmiranda. (#2689)PR: #2638 (@pfistl)
Added several learners for regression and classification on functional data
Added preprocessing steps for feature extraction from functional data
Fixed a bug where multiclass to binaryclass reduction techniques did not work with functional data.
Several other minor bug fixes and code improvements
Extended and clarified documentation for several fda components.
tree_method (@albersonmiranda, #2701)filterFeatures(): Arg thresh was not working correctly when applied to ensemble filters. (@annette987, #2699)classif.xgboost which prevented passing a watchlist for binary tasks. This was caused by a suboptimal internal label inversion approach. Thanks to @001ben for reporting (#32) (@mllg)fda.usc learners to work with package version >=2.0glmnet learners to upstream package version 3.0.0xgboost learners to upstream version 0.90.2 (@pat-s & @be-marc, #2681)classif.gbm and regr.gbm. Specifically, param shrinkage now defaults to 0.1 instead of 0.001. Also more choices for param distribution have been added. Internal parallelization by the package is now suppressed (param n.cores). (@pat-s, #2651)h2o.deeplearning learners (@albersonmiranda, #2668)configureMlr() to .onLoad(), possibly fixing some edge cases (#2585) (@pat-s, #2637)h2o.gbm learners were not running until wcol was passed somehow due to an internal bug. In addition, this bug caused another issue during prediction where the prediction data.frame was somehow formatted as a character rather a numeric. Thanks to @nagdevAmruthnath for bringing this up in #2630.method = "vh" for filter randomForestSRC_var.select and return informative error message for not supported values. Also argument conservative can now be passed. See #2646 and #2639 for more information (@pat-s, #2649)praznik_CMIM does no longer return a result for logical features. See https://gitlab.com/mbq/praznik/issues/19 for more informationdata.frame filter values are now returned in a long (tidy) tibble. This makes it easier to apply post-processing methods (like group_by(), etc) (@pat-s, #2456)benchmark() does not store the tuning results ($extract slot) anymore by default.
If you want to keep this slot (e.g. for post tuning analysis), set keep.extract = TRUE.
This change originated from the fact that the size of BenchmarkResult objects with extensive tuning got very large (~ GB) which can cause memory problems during runtime if multiple benchmark() calls are executed on HPCs.benchmark() does not store the created models ($models slot) anymore by default.
The reason is the same as for the $extract slot above.
Storing can be enabled using models = TRUE.generateFeatureImportanceData() gains argument show.info which shows the name of the current feature being calculated, its index in the queue and the elapsed time for each feature (@pat-s, #26222)classif.liquidSVM and regr.liquidSVM have been removed because liquidSVM has been removed from CRAN.data.tables default in rbindlist(). See #2578 for more information. (@mllg, #2579)regr.randomForest gains three new methods to estimate the standard error:
se.method = "jackknife"
se.method = "bootstrap"
se.method = "sd"?regr.randomForest for more details.regr.ranger relies on the functions provided by the package ("jackknife" and "infjackknife" (default))regr.gbm now supports quantile distribution (@bthieurmel, #2603)classif.plsdaCaret now supports multiclass classification (@GegznaV, #2621)getClassWeightParam() now also works for Wrapper* Models and ensemble models (@ja-thomas, #891)getLearnerNote() to query the "Note" slot of a learner (@alona-sydorova, #2086)e1071::svm() now only uses the formula interface if factors are present. This change is supposed to prevent from "stack overflow" issues some users encountered when using large datasets. See #1738 for more information. (@mb706, #1740)cluster.MiniBatchKmeans from package ClusterR (@Prasiddhi, #2554)plotHyperParsEffect() now supports facet visualization of hyperparam effects for nested cv (@MasonGallo, #1653)data.tables default in rbindlist(). See #2578 for more information. (@mllg, #2579)options(on.learner.error) was not respected in benchmark(). This caused benchmark() to stop even if it should have continued including FailureModels in the result (@dagola, #1984)getClassWeightParam() now also works for Wrapper* Models and ensemble models (@ja-thomas, #891)getLearnerNote() to query the "Note" slot of a learner (@alona-sydorova, #2086)praznik_mrmr also supports regr and surv tasksplotFilterValues() got a bit "smarter" and easier now regarding the ordering of multiple facets. (@pat-s, #2456)filterFeatures(), generateFilterValuesData() and makeFilterWrapper() gained new examples. (@pat-s, #2456)makeResampleDesc(fixed = TRUE)) (@pat-s, #2412).Task help pages are now split into separate ones, e.g. RegrTask, ClassifTask (@pat-s, #2564)deleteCacheDir(): Clear the default mlr cache directory (@pat-s, #2463)getCacheDir(): Return the default mlr cache directory (@pat-s, #2463)getResamplingIndices(inner = TRUE) now correctly returns the inner indices (before inner indices referred to the subset of the respective outer level train set) (@pat-s, #2413).fw.perc, fw.abs or fw.threshold.
It can be triggered with the new cache argument in makeFilterWrapper() or filterFeatures() (@pat-s, #2463).Additionally, filter names have been harmonized using the following scheme:
Added filters FSelectorRcpp_gain.ratio, FSelectorRcpp_information.gain and FSelectorRcpp_symmetrical.uncertainty from package FSelectorRcpp.
These filters are ~ 100 times faster than the implementation of the FSelector pkg.
Please note that both implementations do things slightly different internally and the FSelectorRcpp methods should not be seen as direct replacement for the FSelector pkg.
filter names have been harmonized using the following scheme:
information.gain -> FSelector_information.gain
gain.ratio -> FSelector_gain.ratio
symmetrical.uncertainty -> FSelector_symmetrical.uncertainty
chi.squared -> FSelector_chi.squared
relief -> FSelector_relief
oneR -> FSelector_oneR
randomForestSRC.rfsrc -> randomForestSRC_importance
randomForestSRC.var.select -> randomForestSRC_var.select
randomForest.importance -> randomForest_importance
fixed a bug related to the loading of namespaces for required filter packages (@pat-s, #2483)
"h2o.use.data.table" = TRUE is now the default (@j-hartshorn, #2508)x.bit.names that stores the optimal bitsx now always contains the real feature names and not the bit.namesmakeFeatSelWrapper usable with custom bit.names.sffs crashed in some cases (@bmihaljevic, #2486)