NEWS.md

In NorskRegnesentral/shapr: Prediction Explanation with Dependence-Aware Shapley Values

shapr (development version)

• Release a Python wrapper (shaprpyr, #325) for explaining predictions from Python models (from Python) utilizing almost all functionality of shapr. The wrapper moves back and forth back and forth between Python and R, doing the prediction in Python, and almost everything else in R. This simplifies maintenance of shaprpy significantly. The wrapper is available here.
• Complete restructuring motivated by introducing the Python wrapper. The restructuring splits the explanation tasks into smaller pieces, which was necessary to allow the Python wrapper to move back and forth between R and Python.
• As part of the restructuring, we also did a number of design changes, resulting in a series of breaking changes described below.

Breaking changes

• Moved from explaining predictions using two functions (shapr() for initial setup + explain() for explanation for specific observations), to a single function call (also named explain()). The data used for training and to be explained have gotten explicit names (x_train and x_explain). The order of the input arguments has also been slightly changed (model is now the first argument).
• Prediction and checking functions for custom models are now passed directly as arguments to explain() instead of being defined as functions of a specific class in the global env.
• The previously exported function make_dummies used to explain xgboost models with categorical data, is removed to simplify the code base. This is rather handled with a custom prediction model.
• The function explain.ctree_comb_mincrit, which allowed combining models with approch=ctree with different mincrit parameters, has been removed to simplify the code base. It may return in a completely general manner in later version of shapr.

New features

• Introduce batch computation of conditional expectations (#244). This essentially compute $v(S)$ for a portion of the $S$-subsets at a time, to reduce the amount of data needed to be held in memory. The user can control the number of batches herself, but we set a reasonable value by default (#327). This allows models with large number of features to be explained with a significantly lower RAM consumption (at the cost of a slight increase in the computation time)
• Parallelization over batches (#38) using the future framework.
• Progress bar (#293) using the progressr package. Must be activated by the user with progressr::handlers(global = TRUE) or wrapping the call to explain() around progressr::with_progress({})
• Added approach = 'categorical' (#256, #307) used to explain models with solely categorical features by directly using/estimating the joint distribution of all feature combinations.
• Added approch='timeseries' (#314) for explaining classifications based on time series data/models with the method described in Sec 4.3 of the groupShapley paper.
• Implemented unique sampling of Shapley value subsets (#227)
• Added new function explain_forecast to explain forecasts from time series models, at various prediction horizons (#328). Uses a different set of input argument which is more appropriate for these models.
• Re-implementation of approach = 'independence' method providing significantly faster computation (no longer as a special case of the empirical method). Also allow the method to be used on models with categorical data (#315).
• Added 'beeswarm' and 'waterfall' plots + new coloring scheme for all plots. See the vignette for examples.
• Added timing of the various parts of the explanation process.

Under the hood

• The test base have been completely rewritten (#249). Now heavily utilizing snapshots on a large set of benchmark calls to explain, also using vdiffr for plot tests. Test functions are only written for exported core functions. Internal functions are only tested through the exported ones.
• Update GitHub actions (#335).
• Avoid unnecessary computation of inverse for weight matrix (#280)

Minor improvements and bug fixes

• The vignette/readme/tests now uses the datasets::airquality dataset. This avoids including a new package just for the dataset (#248).
• Allows lm/glm/gam models with interactions (#303). Previously, this was not possible with the prediction functions defined internally due to a bug.
• Sampling of group subsets implemented also for grouping, not only features.

Documentation improvements

• The vignette has been updated to reflect the new framework for explaining predictions, and all the new package features/functionality.

shapr 0.2.3 (GitHub only)

• Development version
• Added support for groupSHAP, including check of appropriate groups, examples and tests
• Various modifications to input of internal functions to reflect that Shapley values may be computed both feature-wise and group-wise
• Fixed bug when passing non-named data to shapr() or explain() (e.g. shapr(data[,1:5],model...)

shapr 0.2.2

• Patch to fix failing CRAN-tests on R-devel due to changed behavior of attach(): Fixed by changing how we simluate adding a function to .GlobalEnv in the failing test. Actual package not affected.

shapr 0.2.1

• Patch to fix warning from development version of data.table due to the use of nomatch argument in merge(), as requested from data.table developers.

shapr 0.2.0

• Minor CRAN release
• Added the new dependence modeling approach "ctree" which handles categorical features in addition to numerical ones. For more information see our paper https://doi.org/10.1007/978-3-030-57321-8_7
• Added support to explain models which take as input categorical features for model classes like xgboost which originally takes only numeric input. On the user side, an additional call to the new make_dummies function is required. See the vignette for details.
• Slight change in the user procedure for explaining predictions from custom models. This now requires only a single function predict_model.
• Introduced a thorough system for extracting and checking the feature information in the model and the data passed to shapr and explain. The features in the data are checked for consistency with what can be extracted from the model object. If the model object is missing some of the necessary information, the info from the data is used instead. The system checks feature labels, classes, and any factor levels.
• Due to the previous point, the feature_names option previously used for custom models is removed.
• Added a manual testing script for custom model (currently cannot be handled by testthat due to environment issues).
• A few under-the-hood changes for checking in the shapr function.

shapr 0.1.4

• Patch to fulfill CRAN policy of using packages under Suggests conditionally (in tests and examples)

shapr 0.1.3

• Fix installation error on Solaris
• Updated README with CRAN installation instructions and badges

shapr 0.1.2

• CRAN release
• Removed unused clustering code
• Removed several package dependencies
• Moved automatic check and pkgdown site build from Circle CI to GitHub actions
• Some minor efficiency fixes
• Changed stopping threshold from 12 to 13 features for none-sampling version of KernelSHAP for consistency with our recommendation
• Changed package title (shortened)
• Minor fixes to fulfill CRAN policy
• Improved documentation
• Revised internal/external and exported/non-exported functions, leading to far fewer external functions and a cleaner manual.

shapr 0.1.1

• Journal of Open Source Software release
• Improved installation instructions and community guidelines in README
• Improved documentation
• Some minor bugfixes

shapr 0.1.0

• Support for custom models
• Improved documentation
• Automated testing using testthat
• Added vignette that gives an introduction to the package
• Added webpage for package using pkgdown
• Improved API for end user
• Various bugfixes

shapr 0.0.0.9000

• First version of the package. Currently under development.

NorskRegnesentral/shapr documentation built on April 19, 2024, 1:19 p.m.