Changelog

0.1.2 (04/09/2022)

The following bugs are fixed in this release:

• Segmentation:

  • A Segmentation object holds incorrect segment count after manipulation (#39).

  • Slic segmentation fails quietly by not starting the segment count at 1. (This issue appears to have been fixed in scikit-image 0.19.2 and higher.)

• Occlusion:

  • occlude_segments_vectorised (Occlusion) returns an occlusion of incorrect shape if the input array is 2D with just one row (#40).

0.1.1 (10/04/2022)

The following functionality is made available with this release:

	Fairness	Accountability	Transparency
Data/ Features
Models			Submodular Pick
Predictions			Image bLIMEy (LIME-equivalent)

This update focuses on surrogate image explainers for predictions of crisp and probabilistic black-box classifiers. In particular, it implements:

• Segmentation:

  • Segmentation abstract class – Segmentation.

  • Slic segmentation – Slic.

  • QuickShift segmentation – QuickShift.

• Occlusion:

  • Generic image occlusion – Occlusion.

• Sampling:

  • Binary random sampling – random_binary_sampler.

• Incremental model processing:

  • Batch-processing and -transforming data for predicting it with a model – batch_data.

• Surrogate image explainability:

  • bLIMEy-based LIME surrogate image explainer – ImageBlimeyLime.

• Aggregation-based model explainability:

  • Submodular pick – submodular_pick.

Additionally, this release moves away from Travis CI in favour of GitHub Actions.

0.1.0 (18/05/2020)

The following functionality is made available with this release:

	Fairness	Accountability	Transparency
Data/ Features
Models
Predictions			Tabular bLIMEy for regression

This is an incremental update focused on surrogate explainers for black-box regression:

• Surrogate explainers – SurrogateTabularExplainer, TabularBlimeyLime and TabularBlimeyTree – now support black-box regression.

• TabularBlimeyLime now uses the correct feature selection approach.

• The surrogate explanation plotting function – plot_lime – has been cleaned.

• 2 new feature selection approaches have been implemented: highest_weights and forward_selection.

• LIME wrapper has been removed.

• Compatibility with scikit-learn newer than 0.21.x has been added.

This release coincides with publication of a paper describing FAT Forensic in The Journal of Open Source Software (JOSS).

0.0.2 (04/11/2019)

The following functionality is made available with this release:

	Fairness	Accountability	Transparency
Data/ Features
Models
Predictions			Tabular bLIMEy

Included tutorials:

• Explaining Machine Learning Predictions: LIME and Counterfactuals (updated to use fatf.transparency.predictions.surrogate_explainers.TabularBlimeyLime instead of fatf.transparency.predictions.lime.Lime).

Included how-to guides:

• How to build LIME yourself (bLIMEy) – Surrogate Tabular Explainers.

Included code examples:

• Using LIME Explainer (updated to use fatf.transparency.predictions.surrogate_explainers.TabularBlimeyLime instead of fatf.transparency.predictions.lime.Lime) and

• Using a Surrogate Tree Explainer.

bLIMEy

This release adds support for custom surrogate explainers of tabular data called bLIMEy. The two pre-made classes are available as part of the fatf.transparency.predictions.surrogate_explainers module:

• fatf.transparency.predictions.surrogate_explainers.TabularBlimeyTree and

• fatf.transparency.predictions.surrogate_explainers.TabularBlimeyLime.

Since the latter class implements LIME from components available in FAT Forensics, the LIME wrapper available under fatf.transparency.lime.Lime will be retired in release 0.0.3.

To facilitate building custom tabular surrogate explainers a range of functionality has been implemented including: data discretisation, data transformation, data augmentation, data point augmentation, distance kernelisation, scikit-learn model tools, feature selection and surrogate model evaluation.

Other Functionality

Seeding of the random number generators via the fatf.setup_random_seed function can now be done by passing a parameter to this function (in addition to using the FATF_SEED system variable).

0.0.1 (01/08/2019)

This is the initial releases of the package. The following functionality is made available with this release:

	Fairness	Accountability	Transparency
Data/ Features	Systemic Bias (disparate treatment labelling) Sample size disparity (e.g., class imbalance)	Sampling bias Data Density Checker	Data description
Models	Group-based fairness (disparate impact)	Systematic performance bias	Partial dependence Individual conditional expectation
Predictions	Counterfactual fairness (disparate treatment)		Counterfactuals Tabular LIME (wrapper)