NEWS.md

ale (development version)

Breaking changes

• We have deeply rethought how best to structure the objects for this package. As a result, the underlying algorithm for calculating ALE has been completely rewritten to be more scalable.
• In addition to rewriting the code under the hood, the structure of all ale objects has been completely rewritten. The latest objects are not compatible with earlier versions. However, the new structure supports the roadmap of future functionality, so we hope that there will be minimal changes in the future that interrupt backward compatibility.
• We have created several S3 objects to represent different kinds of ale package objects:
  • ale: the core ale package object that holds the results of the [ale()] function.
  • ale_boot: results of the [model_bootstrap()] function.
  • ale_p: p-value distribution information as the result of the [create_p_dist()] function.
• With the extensive rewrite, we no longer depend on {ALEPlot} code and so now claim full authorship of the code. One of the most significant implications of this is that we have decided to change the package license from the GPL 2 to MIT, which permits maximum dissemination of our algorithms.
• Renamed the rug_sample_size argument of ale() to sample_size. Now it reflects the size of data that should be sampled in the ale object, which can be used not only for rug plots but for other purposes.
• [ale_ixn()] has been eliminated and now both 1D and 2D ALE are calculated with the [ale()] function.
• [ale()] no longer produces plots. ALE plots are now created as ale_plot objects that create all possible plots from the ALE data from ale or ale_boot objects. Thus, serializing ale objects now avoids the problems of environment bloat of the included ggplot objects.

Bug fixes

• Gracefully fails when the input data has missing values.

Other user-visible changes

• Confidence regions for 1D ALE are now reported more compactly.
• With the creation of plot() methods, eliminated the compact_plots to ale().
• print() and plot() methods have been added to the ale_plots object.
• A print() method has been added to the ale object.
• Interactions are now supported between pairs of categorical variables. (Before, only numerical pairs or pairs with one numerical and one categorical were supported.)
• Bootstrapping is now supported for ALE interactions.
• ALE statistics are now supported for interactions, including confidence regions.
• Categorical y outcomes are now supported. The plots, though, only plot one category at a time.
• 'boot_data' is now an output option from ale(). It outputs the ALE values from each bootstrap iteration.
• model_bootstrap() has added various model performance measures that are validated using bootstrap validation with the .632 correction.
• The structure of p_funs has been completely changed; it has now been converted to an object named ale_p and the functions are separated from the object as internal functions. The function create_p_funs() has been renamed create_p_dist().
• create_p_dist() now produces two types of p-values via the p_speed argument: 'approx fast' for relatively faster but only approximate values (the default) or 'precise slow' for very slow but more exact values.
• Character input data is now accepted as a categorical datatype. It is handled the same as unordered factors.

Under the hood

One of the most fundamental changes is not directly visible but affects how some ALE values are calculated. In certain very specific cases, the ALE values are now slightly different from those of the reference ALEPlot package. These are only for non-numerical variables for some prediction types other than predictions scaled on the response variable. (E.g., a binary or categorical variable for a logarithmic prediction not scaled to the same scale as the response variable.) We made this change for two reasons: * We can understand our implementation and its interpretation for these edge cases much better than that of the reference ALEPlot implementation. These cases are not covered at all in the base ALE scientific article and they are poorly documented in the ALEPlot code. We cannot help users to interpret results that we do not understand ourselves. * Our implementation lets us write code that scales smoothly for interactions of arbitrary depth. In contrast, the ALEPlot reference implementation is not scalable: custom code must be written for each type and each degree of interaction. Other than for these edge cases, our implementation continues to give identical results to the reference ALEPlot package.

Other notable changes that might not be readily visible to users:

• Moved all performance metrics to a new dedicated package, {staccuracy}.
• Reduced dependencies by doing more with the {rlang} and {cli} packages. Reduced the imported functions to a minimum.
• Package messages, warnings, and errors now use {cli}.
• Replaced {assertthat} with custom validation functions that adapt some {assertthat} code.
• Use helper.R test files so that some testing objects are available to the loaded package.
• Configured {future} parallelization code to restore original values on exit.
• Configured codes that use a random seed to restore the original system seed on exit.
• Improved memory efficiency of ale_p objects.
• Plotting code updated for compatibility with ggplot2 3.5.

Known issues to be addressed in a future version

• Plots that display categorical outcomes all on one plot are yet to be implemented. For now, each class or category must be plotted at a time.
• Effects plots for interactions have not yet been implemented.

ale 0.3.0

The most significant updates are the addition of p-values for the ALE statistics, the launching of a pkgdown website which will henceforth host the development version of the package, and parallelization of core functions with a resulting performance boost.

Breaking changes

• One of the key goals for the {ale} package is that it would be truly model-agnostic: it should support any R object that can be considered a model, where a model is defined as an object that makes a prediction for each input row of data that it is provided. Towards this goal, we had to adjust the custom predict function to make it more flexible for various kinds of model objects. We are happy that our changes now enable support for tidymodels objects and various survival models (but for now, only those that return single-vector predictions). So, in addition to taking required object and newdata arguments, the custom predict function pred_fun in the ale() function now also requires an argument for type to specify the prediction type, whether it is used or not. This change breaks previous code that used custom predict functions, but it allows ale to analyze many new model types than before. Code that did not require custom predict functions should not be affected by this change. See the updated documentation of the ale() function for details.

• Another change that breaks former code is that the arguments for model_bootstrap() have been modified. Instead of a cumbersome model_call_string, model_bootstrap() now uses the {insight} package to automatically detect many R models and directly manipulate the model object as needed. So, the second argument is now the model object. However, for non-standard models that {insight} cannot automatically parse, a modified model_call_string is still available to assure model-agnostic functionality. Although this change breaks former code that ran model_bootstrap(), we believe that the new function interface is much more user-friendly.

• A slight change that might break some existing code is that the conf_regions output associated with ALE statistics has been restructured. The new structure provides more useful information. See help(ale) for details.

Other user-visible changes

• The package now uses a pkgdown website located at https://tripartio.github.io/ale/. This is where the most recent development features will be documented.
• P-values are now provided for all ALE statistics. However, their calculation is very slow, so they are disabled by default; they must be explicitly requested. When requested, they will be automatically calculated when possible (for standard R model types); if not, some additional steps must be taken for their calculation. See the new create_p_funs() function for details and an example.
• The normalization formula for ALE statistics was changed such that very minor differences from the median are normalized as zero. Before this adjustment, the former normalization formula could give some tiny differences apparently large normalized effects. See the updated documentation in vignette('ale-statistics') for details. The vignette has been expanded with more details on how to properly interpret normalized ALE statistics.
• Normalized ALE range (NALER) is now expressed as percentile points relative to the median (ranging from -50% to +50%) rather than its original formulation as absolute percentiles (ranging from 0 to 100%). See the updated documentation in vignette('ale-statistics') for details.
• Performance has been dramatically improved by the addition of parallelization by default. We use the {furrr} library. In our tests, practically, we typically found speed-ups of n – 2 where n is the number of physical cores (machine learning is generally unable to use logical cores). For example, a computer with 4 physical cores should see at least ×2 speed-up and a computer with 6 physical cores should see at least ×4 speed-up. However, parallelization is tricky with our model-agnostic design. When users work with models that follow standard R conventions, the {ale} package should be able to automatically configure the system for parallelization. But for some non-standard models users may have to explicitly list the model's packages in the new model_packages argument so that each parallel thread can find all necessary functions. This is only a concern if you get weird errors. See help(ale) for details.
• Fully documented the output of the ale() function. See help(ale) for details.
• The median_band_pct argument to ale() now takes a vector of two numbers, one for the inner band and one for the outer.
• Switched recommendation of calculating ALE data on test data to instead calculate it on the full dataset with the final deployment model.
• Replaced {gridExtra} with {patchwork} for examples and vignettes for printing plots.
• Separated ale() function documentation from ale-package documentation.
• When p-values are provided, the ALE effects plot now shows the NALED band instead of the median band.
• alt tags to describe plots for accessibility.
• More accurate rug plots for ALE interaction plots.
• Various minor tweaks to plots.

Under the hood

• Uses the {insight} package to automatically detect y_col and model call objects when possible; this increases the range of automatic model detection of the ale package in general.
• We have switched to using the {progressr} package for progress bars. With the cli progression handler, this enables accurate estimated times of arrival (ETA) for long procedures, even with parallel computing. A message is displayed once per session informing users of how to customize their progress bars. For details, see help(ale), particularly the documentation on progress bars and the silent argument.
• Moved {ggplot2} from a dependency to an import. So, it is no longer automatically loaded with the package.
• More detailed information from internal var_summary() function. In particular, encodes whether the user is using p-values (ALER band) or not (median band).
• Separated validation functions that are reused across other functions to internal validation.R file.
• Added an argument compact_plots to plotting functions to strip plot environments to reduce the size of returned objects. See help(ale) for details.
• Created package_scope environment.
• Many minor bug fixes and improvements. Improved validation of problematic inputs and more informative error messages.
• Various minor performance boosts after profiling and refactoring code.

Known issues to be addressed in a future version

• Bootstrapping is not yet supported for ALE interactions (ale_ixn()).
• ALE statistics are not yet supported for ALE interactions (ale_ixn()).
• ale() does not yet support multi-output model prediction types (e.g., multi-class classification and multi-time survival probabilities).

ale 0.2.0

This version introduces various ALE-based statistics that let ALE be used for statistical inference, not just interpretable machine learning. A dedicated vignette introduces this functionality (see "ALE-based statistics for statistical inference and effect sizes" from the vignettes link on the main CRAN page at https://CRAN.R-project.org/package=ale). We introduce these statistics in detail in a working paper: Okoli, Chitu. 2023. "Statistical Inference Using Machine Learning and Classical Techniques Based on Accumulated Local Effects (ALE)." arXiv. https://doi.org/10.48550/arXiv.2310.09877. Please note that they might be further refined after peer review.

Breaking changes

• We changed the output data structure of the ALE data and plots. This was necessary to add ALE statistics. Unfortunately, this change breaks any code that refers to objects created by the initial 0.1.0 version, especially code for printing plots. However, we felt it was necessary because the new structure makes coding in workflows much easier. See the vignettes and examples for code examples for how to print plots using the new structure.

Other user-visible changes

• We added new ALE-based statistics: ALED and ALER with their normalized versions NALED and NALER. ale() and model_bootstrap() now output these statistics. (ale_ixn() will come later.)
• We added rug plots to numeric values and percentage frequencies to the plots of categories. These indicators give a quick visual indication of the distribution of plotted data.
• We added a vignette that introduces ALE-based statistics, especially effect size measures, and demonstrates how to use them for statistical inference: "ALE-based statistics for statistical inference and effect sizes" (available from the vignettes link on the main CRAN page at https://CRAN.R-project.org/package=ale).
• We added a vignette that compares the ale package with the reference {ALEPlot} package: "Comparison between {ALEPlot} and {ale} packages" (available from the vignettes link on the main CRAN page at https://CRAN.R-project.org/package=ale).
• We added two datasets:
  • var_cars is a modified version of mtcars that features many different types of variables.
  • census is a polished version of the adult income dataset used for a vignette in the {ALEPlot} package.
• Progress bars show the progression of the analysis. They can be disabled by passing silent = TRUE to ale(), ale_ixn(), or model_bootstrap().
• The user can specify a random seed by passing the seed argument to ale(), ale_ixn(), or model_bootstrap().

Under the hood

By far the most extensive changes have been to assure the accuracy and stability of the package from a software engineering perspective. Even though these are not visible to users, they make the package more robust with hopefully fewer bugs. Indeed, the extensive data validation may help users debug their own errors.

• Added data validation to exported functions. Under the hood, each user-facing function carefully validates that the user has entered valid data using the {assertthat} package; if not, the function fails quickly with an appropriate error message.
• Created unit tests for exported functions. Under the hood, the {testthat} package is now used for testing the outputs of each user-facing function. This should help the code base to be more robust going forward with future developments.
• Most importantly, we created tests that compare results with the original reference {ALEPlot} package. These tests should ensure that any future code that breaks the accuracy of ALE calculations will be caught quickly.
• Bootstrapped ALE values are now centred on the mean by default, instead of on the median. Mean averaging is generally more stable, especially for smaller datasets.
• The code base has been extensively reorganized for more efficient development moving forward.
• Numerous bugs have been fixed following internal usage and testing.

Known issues to be addressed in a future version

• Bootstrapping is not yet supported for ALE interactions (ale_ixn()).
• ALE statistics are not yet supported for ALE interactions (ale_ixn()).

ale 0.1.0

This is the first CRAN release of the ale package. Here is its official description with the initial release:

Accumulated Local Effects (ALE) were initially developed as a model-agnostic approach for global explanations of the results of black-box machine learning algorithms. (Apley, Daniel W., and Jingyu Zhu. "Visualizing the effects of predictor variables in black box supervised learning models." Journal of the Royal Statistical Society Series B: Statistical Methodology 82.4 (2020): 1059-1086 doi:10.1111/rssb.12377.) ALE has two primary advantages over other approaches like partial dependency plots (PDP) and SHapley Additive exPlanations (SHAP): its values are not affected by the presence of interactions among variables in a model and its computation is relatively rapid. This package rewrites the original code from the 'ALEPlot' package for calculating ALE data and it completely reimplements the plotting of ALE values.

(This package uses the same GPL-2 license as the {ALEPlot} package.)

This initial release replicates the full functionality of the {ALEPlot} package and a lot more. It currently presents three functions:

• ale(): create data for and plot one-way ALE (single variables). ALE values may be bootstrapped.
• ale_ixn(): create data for and plot two-way ALE interactions. Bootstrapping of the interaction ALE values has not yet been implemented.
• model_bootstrap(): bootstrap an entire model, not just the ALE values. This function returns the bootstrapped model statistics and coefficients as well as the bootstrapped ALE values. This is the appropriate approach for small samples.

This release provides more details in the following vignettes (they are all available from the vignettes link on the main CRAN page at https://CRAN.R-project.org/package=ale):

• Introduction to the ale package
• Analyzing small datasets (fewer than 2000 rows) with ALE
• ale() function handling of various datatypes for x