effectplots 0.2.0

Major bug fixes

• The outlier clipping algorithm has unintentionally modified the values in place, i.e., also in the original dataframe. This is fixed by #24.

Efficiency improvements

• Significant speed-up and memory reduction for numeric features #16, #24, #25.
• The barebone ALE function .ale() has become faster thanks to issue #11 by @SebKrantz.
• Subsampling indices for outlier capping is now done only once, instead of once per feature #15.

Minor bug fixes

• NA values in feature columns have not been counted in the counts “N”.
• Ordered factors are now working properly.
• ALE are correct also with empty bins at the border (could happen with user-defined breaks).
• update(collapse_m = ...) has collapsed wrong categories #31, #34, and #35.

Documentation

• README has received examples for Tidymodels and probabilistic classification.
• Updated function documentation #41.

Other changes

• Plots with more than one line now use “Effect” als default y label.
• Automatic break count selection via “FD”, “Scott” and via function is not possible anymore #24.
• Export of fcut(), a fast variant of cut() #25.
• x axes are not collected anymore by {patchwork} #27.
• The default of discrete_m = 5 has been increased to 13 #29.
• Slightly different check/preparation of predictions (and the argument pred). Helps to simplify the use of {h2o} #32.
• Updated Plotly subplots layout #33, #43, #44, #45.
• Better test coverage, e.g., #34.
• (Slowish) support for h2o models #36.
• Row names of statistics of numeric features are now removed #37.
• ALE values are now plotted at the right bin break (instead of bin mean) #38.
• Empty factor levels in features are not anymore dropped. However, you can use update(..., drop_empty = TRUE) to drop them after calculations #40.
• Better input checks for average_observed(), average_predicted(), and bias() #41.
• plot(): Renamed argument num_points to continuous_points and cat_lines to discrete_lines #42.
• update(): New argument to_factor to turn discrete non-factors to factors #42.
• EffectData class: Discrete feature values in the output class are represented by their original data types instead of converting them to factors #42.
• EffectData class: The data.frames in the output now contain an attributes discrete to distinguish continuous from discrete features #42.
• effect_importance() will produce an error when sorting on non-existent statistic #45.

effectplots 0.1.0

Initial release.