Changelog • mixgb

mixgb 1.5.2

Significantly faster imputation by optimising data preprocessing and the use of RcppArmadillo
Visual diagnostic functions have been move to the vismi package
Documentation

Set drop.unused.levels = FALSE in fac2sparse() to prevent dropping unused levels in factor or ordinal factor.
- Ensure the feature matrix has the same number of columns to feed in XGBoost

Optimised mixgb() for large datasets:
- Significantly faster imputation by optimising data preprocessing and the use of RcppArmadillo
- Enhanced memory efficiency with in-place modifications using data.table
- Bootstrapping option removed from mixgb(). Users can still use bootstrap in the archived function mixgb0().
- PMM is now set to NULL by default.

Improves package documentation regarding the import of xgb.save() and xgb.load() from XGBoost.

Makes the package compatible with XGBoost 2.0.0 with GPU support:
- Introduces a new parameter device.
- Deprecates parameters gpu_id and predictor.
- Sets tree_method = "hist" by default, aligning with XGBoost 2.0.0.

Introduces support for saving imputation models to a local directory through the parameter save.models.folder in mixgb().
- Models save in JSON format using xgb.save(), a method recommended by XGBoost for future compatibility.
- When save.models.folder is specified, the return object of mixgb() includes the current imputed datasets, directories for imputation models, and relevant parameters. This object can save using saveRDS() as it doesn’t directly contain the models. Users can later load this object into R and employ impute_new(object, newdata, ...) for new data imputation.

Introduces default_params(), an auxiliary function for mixgb(), to validate the list of XGBoost hyperparameters supplied by the user. It simplifies hyperparameter modifications without requiring explicit specification of all default values.

Addresses issues related to PMM for multiclass variables.
Updates plot_hist() and plot_bar() to align with changes in ggplot2 3.4.0:
- Replaces ..density.. with after_stat(density) in plot_hist().
- Replaces ..prop.. with after_stat(prop) in plot_bar().

CRAN release: 2023-02-16

Transitions from bootstrapping to subsampling. Subsampling, set at subsample = 0.7, becomes the default method due to identified biases with bootstrapping in certain scenarios.
- Default for mixgb():
  - Subsampling: subsample = 0.7.
  - No bootstrapping: bootstrap = FALSE.

Resolves a minor issue in the createNA() function.
Modifies default settings in mixgb():
- ordinalAsInteger: Changes from TRUE to FALSE.
- max_depth: Changes from 6 to 3.
- nrounds: Changes from 50 to 100.
- bootstrap: Sets to TRUE by default.

CRAN release: 2022-06-07

Mac OSX users might face challenges with multithreading in mixgb as xgboost requires OpenMP for multi-core operations. For details, please refer to OpenMP for Mac.