Changelog

Source: NEWS.md

effectplots 0.2.3

  • More unit tests.
  • Better code explanations for .ale().

effectplots 0.2.22025-03-09

Minor improvement

  • update(, collapse_m = m) will now produce a level “other p” for the p least frequent categories #52.
  • update() has received new default: collapse_m = 15 (was 30) #54.
  • Missing values in y are now checked after removing observations without positive weights #56.
  • Plotted lines now skip missing values on the y-axis, also in the ribbons #57.
  • Factors with explicit NA level are respected, #59.
  • Empty factor levels are being dropped, #60.

Maintenance

  • Improved logic to find discrete grids for PDPs, #61.
  • Better test coverage for fcut(), #63.

Minor changes

effectplots 0.2.12025-01-11

Major improvement

  • NA values on x axis are always plotted, even for numeric features #49.

Minor changes

  • ggplot plots with no y variation would not show the exposure bars. This has been fixed in #49.
  • Added {labeling} and {scales} explicitly to list of dependencies. Both are required by {ggplot2} anyway #49.

effectplots 0.2.02024-12-11

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 0.1.02024-11-18

Initial release.