• Allow training_frac equal to 1 (#347, @kelly-sovacool).
• Fix for upcoming ggplot2 release (#352, @teunbrand).

• Fix roxygen package doc syntax (r-lib/roxygen2#1491, @kelly-sovacool).

• New functions:
  • bootstrap_performance() allows you to calculate confidence
    intervals for the model performance from a single train/test split by
    bootstrapping the test set (#329, @kelly-sovacool).
  • calc_balanced_precision() allows you to calculate balanced
    precision and balanced area under the precision-recall curve
    (#333, @kelly-sovacool).
• Improved output from find_feature_importance() (#326, @kelly-sovacool).
  • Renamed the column names to feat to represent each feature or group of correlated features.
  • New column lower and upper to report the bounds of the empirical 95% confidence interval from the permutation test.
    See vignette('parallel') for an example of plotting feature importance with confidence intervals.
• Minor documentation improvements (#323, #332, @kelly-sovacool).

Full Changelog: v1.5.0...v1.6.0

• New example showing how to plot feature importances in the parallel vignette (#310, @kelly-sovacool).
• You can now use parRF, a parallel implementation of the rf method, with
  the same default hyperparameters as rf set automatically (#306, @kelly-sovacool).
• New functions to calculate and plot ROC and PRC curves: (#321, @kelly-sovacool)
  • calc_model_sensspec() - calculate sensitivity, specificity, and precision for a model.
  • calc_mean_roc() & plot_mean_roc() - calculate & plot specificity and mean sensitivity for multiple models.
  • calc_mean_prc() & plot_mean_prc() - calculate & plot recall and mean precision for multiple models.

Full Changelog: v1.4.0...v1.5.0

• Extra arguments given to run_ml() are now forwarded to caret::train() (#304, @kelly-sovacool).
  • Users can now pass any model-specific arguments (e.g. weights) to caret::train(), allowing greater flexibility.
• Improved tests (#298, #300, #303 #kelly-sovacool)
• Minor documentation improvements.

Full Changelog: v1.3.0...v1.4.0

• mikropml now requires R version 4.1.0 or greater due to an update in the randomForest package (#292).
• New function compare_models() compares the performance of two models with a permutation test (#295, @courtneyarmour).
• Fixed a bug where cv_times did not affect the reported repeats for cross-validation (#291, @kelly-sovacool).
• Made minor documentation improvements (#293, @kelly-sovacool)

Full Changelog: v1.2.2...v1.3.0

This minor patch fixes a test failure on platforms with no long doubles. The actual package code remains unchanged.

Full Changelog: v1.2.1...v1.2.2

• Allow kfold >= length(groups) (#285, @kelly-sovacool).
  • When using the groups parameter, groups are kept together in cross-validation partitions when kfold <= the number of groups in the training set. Previously, an error was thrown if this condition was not met. Now, if there are not enough groups in the training set for groups to be kept together during CV, groups are allowed to be split up across CV partitions.
• Report p-values for permutation feature importance (#288, @kelly-sovacool).

Full Changelog: v1.2.0...v1.2.1

• New parameter cross_val added to run_ml() allows users to define their own custom cross-validation scheme (#278, @kelly-sovacool).
  • Also added a new parameter calculate_performance, which controls whether performance metrics are calculated (default: TRUE). Users may wish to skip performance calculations when training models with no cross-validation.
• New parameter group_partitions added to run_ml() allows users to control which groups should go to which partition of the train/test split (#281, @kelly-sovacool).
• Modified the training_frac parameter in run_ml() (#281, @kelly-sovacool).
  • By default, training_frac is a fraction between 0 and 1 that specifies how much of the dataset should be used in the training fraction of the train/test split.
  • Users can instead give training_frac a vector of indices that correspond to which rows of the dataset should go in the training fraction of the train/test split. This gives users direct control over exactly which observations are in the training fraction if desired.

• Fixed bugs related to grouping correlated features (#276, @kelly-sovacool).
  • Also, group_correlated_features() is now a user-facing function.