Skip to content

NEWS.md

Latest commit

 

History

History
739 lines (578 loc) · 34.8 KB

NEWS.md

File metadata and controls

739 lines (578 loc) · 34.8 KB
Raw

MachineShop News

Version Updates

3.9.2

  • Remove dead Rd links to the ParBayesianOptimization package.
  • Set required java parameters in the BARTMachineModel example.

3.9.1

  • Ensure that global settings are applied to compute nodes when varimp() runs in parallel.
  • Compatibility fixes for XGBModel (XGBDARTModel, XGBLinearModel, XGBTreeModel).

3.9.0

  • Add offset support to XGBModel.
  • Add logical argument pool to calibration() indicating whether to compute a single calibration curve on predictions pooled over all resampling iterations or to compute them for each iteration individually and return the mean calibration curve.
    • The new argument default is pool = FALSE. The pooling that had been the only implementation in previous package versions (<= 3.8.0) can reproduced with pool = FALSE but is deprecated and will be removed along with the argument in a future version.
    • Note that pooling can result in large memory allocation errors when fitting smooth curves with breaks = NULL.

3.8.0

  • Changes to varimp() arguments.
    • Add argument sort.
    • Extend argument scale to vectors of logical.
  • Changes to model-based variable importance.
    • Fix unused argument error from CForestModel.
    • Use drop1() to compute model term-specific p-values for CoxModel, POLRModel, and SurvRegModel as is done for GLMModel and LMModel.
  • Changes to VariableImportance class.
    • Add slots method and metric to store the computational method ("permute" or "model") and the performance metric used for computations.
    • Add update() method to add the new slots to objects created with previous versions of the package.
  • Deprecate type = "default" option in predict() and replace it with type = "raw".
  • Fix unimplemented type 'list' in 'listgreater' error from SelectedInput.recipe().

3.7.0

  • Compatibility updates for parsnip.
  • Enable resampling by a grouping variable with BootControl, OOBControl, and SplitControl.
  • Enable resampling by a stratification variable with SplitControl.
  • Require R 4.1.0 or later.

3.6.2

  • Add backward compatibility for older MLModel objects without a na.rm slot.
  • Fix CRAN check warning: S3 generic/method consistency.
  • Update role_binom(), role_case(), and role_surv() to remove the requirement that their variables be present in newdata supplied to predict().

3.6.1

  • Compatibility updates for ggplot2, Matrix, and recipes package dependencies.

3.6.0

  • Add argument na.rm to MLModel() for construction of a model that automatically removes all cases with missing values from model fitting and prediction, none, or only those whose missing values are in the response variable. Set the na.rm values in supplied MLModels to automatically remove cases with missing values if not supported by their model fitting and prediction functions.
  • Add argument prob.model to SVMModel().
  • Add argument verbose to fit() and predict().
  • Fix Error in as.data.frame(x) : object 'x' not found issue when fitting a BARTMachineModel that started occurring with bartMachine package version 1.2.7.
  • Remove expired deprecations of ModeledInput and rpp().
  • Internal changes
    • Add slot na.rm to MLModel.

3.5.0

  • Add argument method to r2() for calculation of Pearson or Spearman correlation.
  • Add predict() S4 method for MLModelFit.
  • Export MLModelFunction().
  • Export as.MLInput() methods for MLModelFit and ModelSpecification.
  • Export as.MLModel() method for ModelSpecification.
  • Improve recursive feature elimination of SelectedInput terms.
  • Improve speed of StackedModel and SuperModel.
  • Internal changes
    • Add .MachineShop list attribute to MLModelFit.
    • Move field mlmodel in MLModelFit to model in .MachineShop.
    • Move slot input in MLModel to .MachineShop.
    • Pass .MachineShop to the predict and varimp slot functions of MLModel.

3.4.3

  • Fix TypeError in dependence() with numeric dummy variables from recipes.
  • Prep ModelRecipe with retain = TRUE for recipe steps that are skipped, for example, when test datasets are created.
  • Add generalized area under performance curves to auc(), pr_auc(), and roc_auc() for multiclass factor responses.

3.4.2

  • Add argument select to rfe().
  • Fix object perf_stats not found in optim().

3.4.1

  • Add argument conf to set_optim_bayes().
  • Enable global grid expansion and tuning of StackedModel and SuperModel in ModelSpecification().

3.4.0

  • Fixes
    • Enable prediction with survival times of 0.
  • Implement class SelectedModelSpecification.
  • Internal changes
    • Deprecate classes ModeledInput, ModeledFrame, and ModeledRecipe.
    • Remove unused class TunedModeledRecipe.
  • Expire deprecations
    • Remove argument fixed from TunedModel().
    • Remove Grid().
  • Rename rpp() to ppr().
  • Replace ModeledInput() with ModelSpecification().
  • Require R >= 4.0.0.
  • Use Olden algorithm for NNetModel model-specific variable importance.

3.3.1

  • Fixes
    • SurvRegModelFit summary() error
    • update number of folds recorded in CVControl when stratification or grouping size leads to construction of fewer than requested folds for cross-validation resampling

3.3.0

  • Add argument .type with options "glance" and "tidy" to summary.MLModelFit().
  • Add case components data (stratification and grouping variables) to print.Resample().
  • Add class and methods for ModelSpecification.
  • Add training parameters set functions
    • set_monitor(): monitoring of resampling and optimization
    • set_optim_bayes(): Bayesian optimization with a Gaussian process model
    • set_optim_bfgs(): low-memory quasi-Newton BFGS optimization
    • set_optim_grid(): exhaustive and random grid searches
    • set_optim_method(): user-defined optimization functions
    • set_optim_pso(): particle swarm optimization
    • set_optim_sann(): simulated annealing
  • Add performance() method for MLModel to replicate the previous behavior of summary.MLModel().
  • Add performance(), plot(), and summary() methods for TrainingStep.
  • Add support for unordered plots of Resample performances.
  • Changes to argument type of predict().
    • Add option "default" for model-specific default predictions.
    • Add option "numeric" for numeric predictions.
    • Change option "prob" to be for probabilities between 0 and 1.
  • Change confusion() default behavior to convert factor probabilities to levels.
  • Rename argument control to object in set functions.
  • Rename argument f to fun in roc_index().
  • Return a ListOf training step summaries from summary.MLModel().
  • Return a TrainingStep object from rfe().
  • Support tibble-convertible objects as arguments to expand_params().
  • Internal changes
    • Add class EnsembleModel.
    • Add classes MLOptimization, GridSearch, NullOptimization, RandomGridSearch, and SequentialOptimization.
    • Add class NullControl.
    • Add slot control to PerformanceCurve.
    • Add slot method to TrainingStep.
    • Add slot optim to TrainingParams.
    • Add slot params to MLInput.
    • Inherit class SelectedModel from EnsembleModel.
    • Inherit class StackedModel from EnsembleModel.
    • Inherit class SuperModel from StackedModel.
    • Rename slot case_comps to vars in Resample.
    • Rename slot grid to log in TrainingStep.
  • Fixes
    • error predicting single factor response in GLMModel
    • 'size(x@performance, 3)' error in print.TrainingStep()
    • 'Unmatched tuning parameters' error in TunedModel()

3.2.1

  • Fix 'data' argument of wrong type error in terms.formula().
  • Require >= 3.1.0 version of cli package.

3.2.0

  • Add argument distr and method to dependence().
  • Add function ParsnipModel() for model specifications (model_spec) from the parsnip package.
  • Add function rfe() for recursive feature elimination.
  • Add method as.MLModel() for model_spec and ModeledInput.
  • Add support for any model specification whose object has an as.MLModel() method.
  • Add support for cross-validation with case groups.
  • Add support for names in argument metric of auc().
  • Change argument method default from "model" to "permute" in varimp().
  • Change class ModelFrame to an S4 class; generally requires explicit conversion to a data frame with as.data.frame() in MLModel fit and predict functions.
  • Change progress bar display from elapsed to estimated completion time.
  • Changes to global settings
    • Rename stat.Trained to stat.TrainingParams.
    • Remove stats.VarImp.
  • Changes to internal classes
    • Add class ParsnipModel.
    • Add class SurvTimes.
    • Add class TrainingParams.
    • Add class union Grid.
    • Add class union Params.
    • Add column name, selected, and metrics to slot grid of TrainingStep class.
    • Add slot grid to TunedInput.
    • Add slot id to MLInput and MLModel classes.
    • Add slot id and name to TrainingStep class.
    • Add slot models to SelectedModel.
    • Remove slot name from MLControl classes.
    • Remove slot selected, values, and metric from TrainingStep class.
    • Remove slot shift from VariableImportance class.
    • Rename class Grid to TuningGrid.
    • Rename class Resamples to Resample.
    • Rename class TrainStep to TrainingStep.
    • Rename class VarImp to VariableImportance.
    • Rename classes of MLControl.
      • MLBootControl → BootControl
      • MLBootOptimismControl → BootOptimismControl
      • MLCVControl → CVControl
      • MLCVOptimismControl → CVOptimismControl
      • MLOOBControl → OOBControl
      • MLSplitControl → SplitControl
      • MLTrainControl → TrainControl
    • Rename column Input and Model to params in slot grid of TrainingStep class.
    • Rename column Resample to Iteration in Resample class
    • Rename slot x to input in MLModel class.
  • Changes to XGBModel
    • Change argument default for nrounds from 1 to 100.
    • Rearrange constructor arguments.
    • Reduce number of tuning grid parameters
      • Include nrounds and max_depth in automated grids for XGBDARTModel and XGBTreeModel.
      • Include nrounds, lambda, and alpha in automated grid for XGBLinearModel.
    • Compute survival probabilities for survival:aft prediction.
    • Change default survival objective from survival:cox to survival:aft.
  • Format and condense printout of objects.
  • Include all computed performance metrics in TrainingStep objects and output.
  • Remove shift from variable importance scaling in varimp().
  • Rename and redefine dispatch (first) arguments in functions.
    • model → object in TunedModel()
    • x → object in expand_model()
    • x → formula/input/model in expand_modelgrid(), fit(), ModelFrame(), resample(), rfe() methods
    • x → formula/object/model in ModeledInput() methods
    • x → object in ParameterGrid() methods
    • x → control in set_monitor(), set_predict(), set_strata()
    • x → object in TunedInput()
  • Rename function Grid() to TuningGrid().
  • Reorder optional arguments in ModelFrame().
  • Save model constructor arguments as the list elements in MLModel params slots.

3.1.0

  • Add argument na.rm to dependence().
  • Add global setting stats.VarImp for summary statistics to compute on permutation-based variable importance.
  • Add permutation-based variable importance to varimp().
  • Sort variable importance by first column only if not scaled.
  • Correct the estimated variances for cross-validation estimators of mean performance difference in t.test.PerformanceDiff().
  • Rename argument metric to type in varimp() functions for BartMachineModel, C50Model, EarthModel, RFSRCModel, and XGBModel.
  • Set argument type default to "nsubsets" in EarthModel varimp().
  • Expand case weighted metrics support.
    • Fix weights used in survival event-specific metrics.
    • Use weights for cross_entropy() numeric method.
    • Use weights for predicted survival probabilities.
  • Fix error with argument f in roc_index() Surv method.

3.0.0

  • Add slot weights to MLModel classes.
  • Allow case weights in LMModel for all response types.
  • Exclude infinite values from calculation of breaks in calibration().
  • Fix invalid max = Inf arguments to print.default().
  • Add support for case weights in performance metrics and curves.
  • Evaluate ModelFrame() arguments strata and weights in data environment.
  • Fix issue introduced in package version 2.9.0 of recipe case weights not being used in model fitting.
  • Add column Weight of case weights to Resamples data frame.
  • Rename values column to get_values in MLModel gridinfo slot.
  • Move global settings resample_progress and resample_verbose to set_monitor() arguments progress and verbose.
  • Move MLControl() arguments strata_breaks, strata_nunique, strata_prop, and strata_size to set_strata() arguments breaks, nunique, prop, and size.
  • Move MLControl() arguments times, distr, and method to set_predict().
  • Export %>% operator.
  • Return case stratification values in the 'strata' slot of Resamples objects.

2.9.0

  • Rename tibble column regular to default in MLModel gridinfo slot.
  • Redefine size and random arguments of ParameterGrid() to match those of Grid().
  • Revise selection of character values in model grids.
    • Select coeflearn values in their defined order instead of at random in AdaBoostModel.
    • Select kernels values in their defined order instead of at random in KNNModel.
    • Add survival splitrule methods in RangerModel.
    • Select splitrule values in their defined order instead of at random in RangerModel.
  • Revise global settings names.
    • Rename max.print to print_max.
    • Rename progress.resample to resample_progress.
    • Rename stat.train to stat.Trained.
    • Rename dist.Surv to distr.SurvMeans.
    • Rename dist.SurvProbs to distr.SurvProbs.
  • Implement customized stratification methods for resampling.
    • Stratify survival data by time within event status by default instead of by event status only.
    • Add strata_breaks, strata_nunique, strata_prop and strata_size arguments to MLControl() constructor.
    • Reduce strata_breaks if numeric quantile bins are below strata_prop and strata_size.
    • Pool smallest factor levels below strata_prop and strata_size iteratively.
    • Pool smallest adjacent ordered levels below strata_prop and strata_size iteratively.
  • Remove deprecated length arguments from Grid() and ParameterGrid().
  • Drop compatibility with deprecated gridinfo functions in MLModel().
  • New and improved survival analysis methods.
    • Add support for counting process survival data.
    • Use model weights in estimation of predicted baseline survival curves.
    • Change censoring curve estimation method from direct to cumulative hazard-based in the brier() metric.
    • Improve computational speed of survival curve estimation.
    • Remove "fleming-harrington" as a choice for the method argument of predict() and for the method.EmpiricalSurv global setting, because it is a special case of the existing (default) "efron" choice and thus not needed.
    • Add "rayleigh" choice for the distr.Surv and distr.SurvProbs global settings.
  • Rename dist argument to distr in calibration(), MLControl(), predict(), and r2().
  • Return survival distribution name with predicted values.
    • Add distr argument to SurvEvents() and SurvProbs().
    • Add SurvMeans class.
    • Return predicted mean survival times as SurvMeans object.
    • Default to the distribution used in predicting mean survival times in calibration() and r2().
  • Rename "terms" predictor_encoding to "model.frame" in MLModel class.
  • Pass elliptical arguments in performance() response type-specific methods to metrics supplied as a single MLMetric function.

2.8.0

  • Replace get_grid() with expand_modelgrid().
  • Fix for truncated grid of lambda values in GLMNetModel.
  • Support package version constraints in MLModel.

2.7.1

  • Rename traininfo slot to train_steps in MLModel classes.
  • Issue #4: compatibility fix for recipes package change in behavior of the retain argument in prep().

2.7.0

  • Sort randomly sampled grid points.
  • Change fixed argument default NULL to list() in TunedModel().
  • CRAN release.

2.6.2

  • Rename length argument to size in Grid() and ParameterGrid().
  • Add support for named sizes in ParameterGrid().
  • Revise model tuning grids.
    • Replace grid slot with gridinfo in MLModel classes.
    • Add support for size vectors in Grid().
    • Add get_grid() function to extract model-defined tuning grids.
  • Rename trainbits slot to traininfo in MLModel classes.

2.6.1

  • Doc edits: do not test examples requiring suggested packages.
  • CRAN release.

2.6.0

  • Preprocess data for automated grid construction only when needed.
  • Select RPartModel cp grid points from cptable according to smallest cross-validation error (mean plus one standard deviation).
  • CRAN release.

2.5.2

  • Export Performance diff() method.

2.5.1

  • Implement fast random forest model RFSRCModel.
  • Export unMLModelFit() function to revert an MLModelFit object to its original class.

2.5.0

  • Add options argument to step_lincomp() and step_sbf().
  • CRAN release.

2.4.3

  • Add recipe step_sbf() function for variable selection by filtering.
  • Inherit step_kmedoids objects from step_sbf, and refactor methods.
    • Support user-specified center and scale functions.
    • Append prefix to selected variable names.
    • Rename tidy() column medoids to selected.
    • Rename tidy() column names to name.
    • Set tidy() non-selected variable names to NA.
  • Add recipe step_lincomp() function for linear components variable reduction.
  • Inherit step_kmeans objects from step_lincomp, and refactor methods.
    • Support user-specified center and scale functions.
    • Rename tidy() column names to name.
  • Inherit step_spca objects from step_lincomp, and refactor methods.
    • Support user-specified center and scale functions.
    • Rename tidy() column value to weight.
    • Rename tidy() column component to name.
  • Set GBMModel distribution to bernoulli, instead of multinomial, for binary responses.

2.4.2

  • Add global setting RHS.formula for listing of operators and functions allowed on right-hand side of traditional formulas.
  • Add clara clustering method to step_kmedoids().
  • Support Cox and accelerated failure time regression for survival responses in XGBModel, XGBDARTModel, XGBLinearModel, and XGBTreeModel.

2.4.1

  • Set NNetModel linout argument automatically according to the response variable type (numeric: TRUE, other: FALSE). Previously, linout had a default value of FALSE as defined in the nnet package.

2.4.0

  • CRAN release.

2.3.2

  • Display progress bars for sequential resampling iterations.

2.3.1

  • R 4.0 data.frame compatibility updates for calibration curves.
  • Fix recipe prediction with StackedModel and SuperModel

2.3.0

  • Display progress messages for any foreach parallel backend.

2.2.5

  • Show all error messages when resample selection stops.
  • Preserve predictor names in NNetModel fit() method.
  • Fix aggregation of performance curves with infinite values.
  • Add progress bar and verbose output options for resample() methods.
  • Get non-negative probabilities for survival confusion matrix.
  • Update Using webpages and vignette.

2.2.4

  • Fix BARTMachineModel to predict highest binary response level.
  • Grid tune BARTMachineModel nu parameter for numeric responses only.

2.2.3

  • Extend ModeledInput() to SelectedModelFrame, SelectedModelRecipe, and TunedModelRecipe.

2.2.2

  • Fix updating of recipe parameters in TunedInput().

2.2.1

  • Print StackedModel and SuperModel training information.
  • Fix missing case names when resampling with recipes.

2.2.0

  • CRAN release.

2.1.4

  • Add cost-complexity pruning parameters to TreeModel.
  • Perform stratified resampling automatically for ModeledInput() and SelectedInput() objects constructed with formulas and matrices.

2.1.3

  • Revisions needed to some fit() methods to ensure that unprepped recipes are passed to models, like TunedModed, StackedModel, SelectedModel and SuperModel, needing to replicate preprocessing steps in their resampling routines.
  • Extend GLMModel to factor and matrix responses.
  • Use fun instead of deprecated fun.y in ggplot2 functions.
  • Capture user-supplied parameters passed in to the ellipsis of model constructor functions that have them.

2.1.2

  • Compatibility fix for tibble 3.0.0.
  • Include missing values in model matrices created internally from formulas.

2.1.1

  • Improve specificity of metricinfo() results for factor responses.
  • Correct SplitControl() to train on the split sample instead of the full dataset.
  • Perform stratified resampling automatically when fit() formula and matrix methods are called with meta-models.

2.1.0

  • CRAN release.

2.0.4

  • Extend print() argument n to data frame and matrix columns for more concise display of large data structures.
  • Add preprocessing recipe functions step_kmeans(), step_kmedoids(), and step_spca().

2.0.3

  • Internal changes:
    • Remove MLModel slot y.
    • Rename ModelFrame and ModelRecipe columns (casenames) to (names).
    • Register ModelFrame inheritance from data.frame.
    • Define Terms S4 classes for ModelFrame slot terms.

2.0.2

  • Implement ModeledInput, SelectedInput and TunedInput classes and methods.
  • Deprecate SelectedFormula(), SelectedMatrix(), SelectedModelFrame(), SelectedRecipe(), and TunedRecipe().
  • Remove deprecated tune().
  • Rename global setting stat.Curves to stat.Curve.

2.0.1

  • Rename global setting stat.Train to stat.train.
  • Add print methods for SelectedModel, StackedModel, SuperModel, and TunedModel.
  • Revise training methods to ensure nested resampling of SelectedRecipe and TunedRecipe.
  • Return list of all training steps in MLModel trainbits slot.

2.0.0

  • Rename global setting stat.Tune to stat.Train.
  • Enable selection of formulas, design matrices, and model frames with SelectedFormula(), SelectedMatrix(), and SelectedModelFrame().
  • Rename discrete variable classes: BinomialMatrix → BinomialVariate, DiscreteVector → DiscreteVariate, NegBinomialVector → NegBinomialVariate, and PoissonVector → PoissonVariate.
  • Add global setting require for user-specified packages to load during parallel execution of resampling algorithms.
  • Rename recipe role case_strata to case_stratum.
  • Rename object argument to data in ConfusionMatrix(), SurvEvents(), and SurvProbs().
  • Add c methods for BinomialVariate, DiscreteVariate, ListOf, and SurvMatrix.
  • Add role_binom(), role_case(), role_surv(), and role_term() to set recipe roles.
  • Support base argument to varimp() for log-transformed p-values.
  • Rename ParamSet to ParameterGrid.
  • Add option to reset global settings individually.
  • Add as.data.frame methods for Performance, Performance summary, PerformanceDiff, PerformanceDiffTest, and Resamples.

1.99.0

  • Implement DiscreteVector class and subclasses BinomialVector, NegBinomialVector, and PoissonVector for discrete response variables.
  • Extend model support to DiscreteVector classes as follows.
    • DiscreteVector: all models applicable to numeric responses.
    • BinomialVector/NegBinomialVector/PoissonVector: BlackBoostModel, GAMBoostModel, GLMBoostModel, GLMModel, and GLMStepAICModel.
    • BinomialVector/PoissonVector: GLMNetModel.
    • PoissonVector: GBMModel and XGBModel
  • Add support for offset terms in formulas, model matrices, and recipes.
  • Add recipe tune information to fitted MLModel.
  • Replace Calibration(), Confusion(), Curves(), Lift(), and Resamples() with c methods.
  • Redefine Confusion S3 class as ConfusionList S4 class.
  • Remove support for one-element list to metricinfo() and modelinfo().
  • Remove deprecated expand.model().
  • Expire deprecated tune().

1.6.4

  • Calculate regression variable importance as negative log p-values.
  • Support empty vectors in metricinfo() and modelinfo().
  • Add support for dials package parameter sets with ParamSet().

1.6.3

  • Add as.MLModel() for coercing MLModelFit to MLModel.
  • Deprecate tune(); call fit() with a SelectedModel or TunedModel instead.

1.6.2

  • Implement optimism-corrected cross-validation (CVOptimismControl).
  • Fix BootOptimismControl error with 2D responses.
  • Add global option max.print for the number of models and data frame rows to show with print methods.
  • Enable recipe selection with SelectedRecipe().
  • Refactor tune() methods.
  • Replace MLModelFit element fitbits (MLFitBits object) with mlmodel (MLModel object).
  • Rename VarImp slot center to shift.

1.6.1

  • Use tibbles for parameter grids.
  • Add random sampling option to expand_model(), expand_params(), and expand_steps().
  • Display information for model functions and objects more compactly.

1.6.0

  • Add global setting for default cutoff threshold value.
  • Add option to reset all global settings.
  • Enable recipe tuning with TunedRecipe().
  • Add expand_model() for model expansion over tuning parameters.
  • Add expand_params() for model parameters expansion.
  • Add expand_steps() for recipe step parameters expansion.
  • Implement MLModelFunction and MLModelList classes.
  • Add fit methods for MLModel, MLModelFunction, and MLModelList.
  • Fix NNetModel fit error with binary and factor responses.
  • Fix modelinfo() function not found error.

1.5.2

  • Implement exception handling of tune() resampling failures.
  • Remove deprecated types and design arguments from MLModel().

1.5.1

  • Implement global settings for default resampling control, performance metrics, summary statistics, and tuning grid.
  • Support vector arguments in metricinfo() and modelinfo().
  • Update package documentation.

1.5.0

  • Implement model: SelectedModel.
  • Remove maximize argument from tune() and TunedModel.
  • Support lists as arguments to StackedModel() and SuperModel.

1.4.2

  • Revert renaming of expand.model().
  • Exclude 0 distance from KNNModel tuning grid.
  • Improve random tuning grid coverage.

1.4.1

  • Implement model: TunedModel.
  • Remove deprecated na.action argument from ModelFrame methods.
  • Rename MLModel() argument types to response_types.
  • Rename MLModel() argument design to predictor_encoding.
  • Rename expand.model() to expand_model().

1.4.0

  • CRAN release.

1.3.3

  • Implement optimism-corrected bootstrap resampling (BootOptimismControl).
  • Store case names in ModelFrame and ModelRecipe and save to Resamples.

1.3.2

  • Add BinaryConfusionMatrix and OrderedConfusionMatrix classes.
  • Export ConfusionMatrix constructor.
  • Extend metricinfo() to confusion matrices.
  • Refactor performance metrics methods code.

1.3.1

  • Check and convert ordered factors in response methods.
  • Check consistency of extracted variables in response methods.
  • Add metrics methods for Resamples.

1.3.0

  • Improve compatibility with preprocessing recipes.
  • Allow base math functions and operators in ModelFrame formulas.

1.2.5

  • Save ModelFrame response in first column.
  • Unexport response formula method.
  • Add ICHomes dataset.
  • Add center and scale slot to VarImp.

1.2.4

  • Prohibit in-line functions in ModelFrame formulas.
  • Rename response function argument from data to newdata.

1.2.3

  • Add fit, resample, and tune methods for design matrices.
  • Reduce computational overhead for design matrices and recipes.
  • Rename ModelFrame() argument na.action to na.rm.

1.2.2

  • Implement parametric ("exponential", "rayleigh", "weibull") estimation of baseline survival functions.
  • Set "weibull" as the default distribution for survival mean estimation.
  • Add extract method for Resamples.
  • Add na.rm argument to calibration(), confusion(), performance(), and performance_curve().
  • Add loess span argument to calibration().
  • Change SurvMatrix from S4 to S3 class.

1.2.1

  • Add method option to predict() for Breslow, Efron (default), or Fleming-Harrington estimation of survival curves for Cox proportional hazards-based models.
  • Add dist option to predict() for exponential or Weibull approximation to estimated survival curves.
  • Add dist option to calibration() for distributional estimation of observed mean survival.
  • Add dist option to r2() for distributional estimation of the total sum of squares mean.
  • Handle unnamed arguments in metricinfo() and modelinfo().

1.2.0

  • Implement metrics: auc, fnr, fpr, rpp, tnr, tpr.
  • Implement performance curves, including ROC and precision recall.
  • Implement SurvMatrix classes for predicted survival events and probabilities to eliminate need for separate times arguments in calibration, confusion, metrics, and performance functions.
  • Add calibration curves for predicted survival means.
  • Add lift curves for predicted survival probabilities.
  • Add recipe support for survival and matrix outcomes.
  • Rename MLControl argument surv_times to times.
  • Fix identification of recipe case_weight and case_strata variables.
  • Launch package website.
  • Bring Introduction vignette up to date with package features.

1.1.0

  • Implement model: BARTModel.
  • Implement model tuning over automatically generated grids of parameter values and random sampling of grid points.
  • Add metrics for predicted survival times: accuracy, f_score, kappa2, npv, ppv, pr_auc, precision, recall, roc_index, sensitivity, specificity
  • Add metrics for predicted survival means: cindex, gini, mae, mse, msle, r2, rmse, rmsle.
  • Add performance and metric methods for ConfusionMatrix.
  • Add confusion matrices for predicted survival times.
  • Standardize predict functions to return mean survival when times are not specified.
  • Replace MLModel slot and constructor argument nvars with design.

1.0.0

  • Implement models: BARTMachineModel, LARSModel.
  • Implement performance metrics: gini, multi-class pr_auc and roc_auc, multivariate rmse, msle, rmsle.
  • Implement smooth calibration curves.
  • Implement MLMetric class for performance metrics.
  • Add as.data.frame method for ModelFrame.
  • Add expand.model function.
  • Add label slot to MLModel.
  • Expand metricinfo/modelinfo support for mixed argument types.
  • Rename calibration argument n to breaks.
  • Rename modelmetrics function to performance.
  • Rename ModelMetrics/Diff classes to Performance/Diff.
  • Change MLModelTune slot resamples to performance.

0.4.0

  • Implement models: AdaBagModel, AdaBoostModel, BlackBoostModel, EarthModel, FDAModel, GAMBoostModel, GLMBoostModel, MDAModel, NaiveBayesModel, PDAModel, RangerModel, RPartModel, TreeModel
  • Implement user-specified performance metrics in modelmetrics function.
  • Implement metrics: accuracy, brier, cindex, cross_entropy, f_score, kappa2, mae, mse, npv, ppv, pr_auc, precision, r2, recall, roc_auc, roc_index, sensitivity, specificity, weighted_kappa2.
  • Add cutoff argument to confusion function.
  • Add modelinfo and metricinfo functions.
  • Add modelmetrics method for Resamples.
  • Add ModelMetrics class with print and summary methods.
  • Add response method for recipe.
  • Export Calibration constructor.
  • Export Confusion constructor.
  • Export Lift constructor.
  • Extend calibration arguments to observed and predicted responses.
  • Extend confusion arguments to observed and predicted responses.
  • Extend lift arguments to observed and predicted responses.
  • Extend metrics and stats function arguments to accept function names.
  • Extend Resamples to arguments with multiple models.
  • Change CoxModel, GLMModel, and SurvRegModel constructor definitions so that model control parameters are specified directly instead of with a separate control argument/structure.
  • Change predict(..., times = numeric()) function calls to survival model fits to return predicted values in the same direction as survival times.
  • Change predict(..., times = numeric()) function calls to CForestModel fits to return predicted means instead of medians.
  • Change tune function argument metrics to be defined in terms of a user-specified metric or metrics.
  • Deprecate MLControl arguments cutoff, cutoff_index, na.rm, and summary.

0.3.0

  • Implement linear models (LMModel), linear discriminant analysis (LDAModel), and quadratic discriminant analysis (QDAModel).
  • Implement confusion matrices.
  • Support matrix response variables.
  • Support user-specified stratification variables for resampling via the strata argument of ModelFrame or the role of "case_strata" for recipe variables.
  • Support user-specified case weights for model fitting via the role of "case_weight" for recipe variables.
  • Provide fallback for models with undefined variable importance.
  • Update the importing of prepper due to its relocation from rsample to recipes.

0.2.0

  • Implement partial dependence, calibration, and lift estimation and plotting.
  • Implement k-nearest neighbors model (KNNModel), stacked regression models (StackedModel), super learner models (SuperModel), and extreme gradient boosting (XGBModel).
  • Implement resampling constructors for training resubstitution (TrainControl) and split training and test sets (SplitControl).
  • Implement ModelFrame class for general model formula and dataset specification.
  • Add multi-class Brier score to modelmetrics().
  • Extend predict() to automatically preprocess recipes and to use training data as the newdata default.
  • Extend tune() to lists of models.
  • Extent summary() argument stats to functions.
  • Fix survival probability calculations in GBMModel and GLMNetModel.
  • Change MLControl argument na.rm default from FALSE to TRUE.
  • Removed na.rm argument from modelmetrics().

0.1

  • Initial public release