ciu-package {ciu}

ciu: Contextual Importance and Utility

Description

Implementation of the Contextual Importance and Utility (CIU) concepts for Explainable AI (XAI). A recent description of CIU can be found in e.g. Främling (2020) arXiv:2009.13996.

Details

This package implements the Contextual Importance and Utility (CIU) concepts for Explainable AI (XAI). CIU allows explaining output values of any regression or classification systems, no matter if it is a "black-box" or a "white-box" AI, or anything between black and white. CIU is entirely model-agnostic. Contrary to most (all?) other XAI methods, CIU provides explanations directly based on the observed input-output behavior without building an intermediate "interpretable" model for doing it.

CIU was developed by Kary Främling in his PhD thesis, which was presented in 1996 (in French). CIU was first presented in 1995 at the International Conference on Artificial Neural Networks (ICANN).

The ciu package supports models from caret and at least lda natively, but can easily be made to work with any model.

Main functions:

Use of ciu starts by calling the function ciu.new that returns an object of class CIU. If the CIU object is created by ciu <- ciu.new(...), then different methods can be called as ciu$explain(), ciu$barplot.ciu() etc. for obtaining explanations in different forms.

ciu is implemented using an "old style" (?) R object orientation. However, it provides object-oriented encapsulation of variables and methods of the CIU object, which presumably helps to avoid name conflicts with other packages or user code.

Since version 0.5.0 it is also possible to use a non-object-oriented approach by creating an ordinary list of class ciu by calling the function ciu. That ciu object is then passed as the first parameter to the different functions. This parallel possibility was originally developed mainly for getting support for proper Roxygen functionality. However, it does also offer some interesting properties, e.g. a CIU object takes up much more memory than a ciu object because it creates its own environment. CIU objects can be converted to ciu objects and vice versa at any time by the ⁠<CIU>$as.ciu()⁠ method and the ciu.to.CIU function.

It is recommended to use the object-oriented approach in order to avoid unnecessary conversions back and forth. However, the difference is presumably not very significant.

References

Främling, K. Explainable AI without Interpretable Model. 2020, https://arxiv.org/abs/2009.13996.

Främling, K. Decision Theory Meets Explainable AI. 2020, <doi.org/10.1007/978-3-030-51924-7_4>.

Främling, K. Modélisation et apprentissage des préférences par réseaux de neurones pour l'aide à la décision multicritère. 1996, https://tel.archives-ouvertes.fr/tel-00825854/document (title translation in English: Learning and Explaining Preferences with Neural Networks for Multiple Criteria Decision Making)