Interactive Explanation and Elicitation for Multiple Criteria Decision Analysis Vincent Mousseau & Wassila Ouerdane Laboratoire Génie Industriel In collaboration with : Kh. Belahcene (LGI), Ch. Labreuche (Thales) and N. Maudet (LIP6) IRT SystemX–April 11, 2018
51
Embed
Interactive Explanation and Elicitation for Multiple ...€¦ · Pairwise comparisons Ordinal Sorting Future prospects and applications 2/44 . Motivations Introduction to Multiple
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Interactive Explanation and Elicitation forMultiple Criteria Decision Analysis
Vincent Mousseau & Wassila Ouerdane
Laboratoire Génie Industriel
In collaboration with : Kh. Belahcene (LGI), Ch. Labreuche (Thales) and
N. Maudet (LIP6)
IRT SystemX–April 11, 2018
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Contents
Motivations
Introduction to Multiple Criteria Decision Aiding
Basic MCDA concepts
Preference Elicitation
Explanation schemes in MCDA context
Pairwise comparisons
Ordinal Sorting
Future prospects and applications
2/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Motivations
I new regulations (eg. GDPR)
I raising concern in the society : making A.I. systems trustable !
Featured in mainstream press, related to prominent applications :
I automated decisions for autonomous vehicles
I loan agreements
I Admission Post Bac/ParcourSup
3/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Motivations
I new regulations (eg. GDPR)
I raising concern in the society : making A.I. systems trustable !
Featured in mainstream press, related to prominent applications :
I automated decisions for autonomous vehicles
I loan agreements
I Admission Post Bac/ParcourSup
3/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Motivations
I new regulations (eg. GDPR)
I raising concern in the society : making A.I. systems trustable !
Featured in mainstream press, related to prominent applications :
I automated decisions for autonomous vehicles
I loan agreements
I Admission Post Bac/ParcourSup
3/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
General Data Protection Regulation : A right to
explanation?
However, in their examination of the legal status of the GDPR, Wachter et
al. conclude that such a right does not exist yet. The right to explanation is
only explicitly stated in a recital :
a person who has been subject to automated decision-making“should be subject to suitable safeguards, which should includespeci�c information to the data subject and the right to obtainhuman intervention, to express his or her point of view, to obtain anexplanation of the decision reached after such assessment and tochallenge the decision ”
However, recitals are not legally binding. It also appears to have been
intentionally not included in the �nal text of the GDPR after appearing in
an earlier draft.
4/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
General Data Protection Regulation : A right to
explanation?
Still, Article 13 and 14 about noti�cation duties may provide a right to be
informed about the “logic involved” prior to decision
“existence of automated decision-making, including pro�ling [...][and provide data subjects with] meaningful information about thelogic involved, as well as the signi�cance and the envisagedconsequences of such processing.”
As it stands, only provides a (limited : secret of a�airs, etc.) right to obtain
ex-ante explanations about the model (which they call, ‘right to be
informed’).
Wachter et al. Why a Right to Explanation of Automated Decision-Making Does Not Exist inthe General Data Protection Regulation. International Data Privacy Law, 2017.
5/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Loi pour une républiqe numériqe
L’administration communique à la personne faisant l’objet d’une décision
individuelle prise sur le fondement d’un traitement algorithmique, à la
demande de celle-ci, sous une forme intelligible et sous réserve de ne pas
porter atteinte à des secrets protégés par la loi, les informations suivantes :
I Le degré et le mode de contribution du traitement algorithmique à la
prise de décision ;
I Les données traitées et leurs sources ;
I Les paramètres de traitement et, le cas échéant, leur pondération,
appliqués à la situation de l’intéressé ;
I Les opérations e�ectuées par le traitement.
Décret du 14 Mars 2017, cité et commenté dans :
Besse et al.. Loyauté des Décisions Algorithmiques. Contribution to CNIL debate, 2017.
6/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Transparency, Interpretability or Explainability?
According to Besse et al., a decision can be said to be :
I transparent when the algorithm/code are made available.
I interpretable when it is possible to identify the features or variables
which were prominent for the decision (even sometimes quantify this
importance)
I explainable when it is possible to explicitly relate the values taken by
the input data and the taken decision
Besse et al.. Loyauté des Décisions Algorithmiques. Contribution to CNIL debate, 2017 (my
translation).
7/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Transparency does not imply explainability
prints Hello World! (by Ben Kurtovic, winner of a 2017 obfuscation contest)
8/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Transparency does not imply explainability
prints Hello World! (by Ben Kurtovic, winner of a 2017 obfuscation contest)8/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
A panel of qestions we need to answer?
1. what were the main factors in a decision?
2. would changing a given factor have changed the decision?
3. how to improve the decision?
4. why did two similar-looking cases get di�erent conclusions, or
vice-versa?
5. does the model indeed do what is expected?
6. why this decision (recommendation) ?
7. ...
9/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
The explanation landscape is rich already
Examples of approaches
I data-based explanations (incl. counterfactuals) [Datta et al., 2016]
I locally faithful approximations (LIME), surrogate models [Ribeiro et
al, 2016]
I add constraints or objective (capturing interpretability) [Sokolovska et
al., 2017] ;
I restrict operators to argumentation schemes validated by the
user. [Belahcène et al., 2017]
I ...
Datta et al.. Algorithmic transparency via quantitative input in�uence : Theory and experi-ments with learning systems. The 37th IEEE Symposium on Security and Privacy.2016.
Ribeiro et al.. “why should i trust you?” Explaining the predictions of any classi�er. In ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining.2016.
Sokolovska et al.. The fused lasso penalty for learning interpretable medical scoring sys-tems.2017. IJCNN.
Belahcène et al.. Explaining robust additive utility models by sequences of preference swaps.Theory and Decision. 2017.
10/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
The explanation landscape is rich already
Examples of approaches
I data-based explanations (incl. counterfactuals) [Datta et al., 2016]
I locally faithful approximations (LIME), surrogate models [Ribeiro et
al, 2016]
I add constraints or objective (capturing interpretability) [Sokolovska et
al., 2017] ;
I restrict operators to argumentation schemes validated by the
user. [Belahcène et al., 2017]
I ...
An explanation (argumentation) scheme
an operator tying a tuple of premises (pieces of information provided or
approved by the Decision Maker, or inferred during the process, and some
supplementary hypotheses on the reasoning process (model’s assumptions)
to a conclusion.
11/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Contents
Motivations
Introduction to Multiple Criteria Decision Aiding
Basic MCDA concepts
Preference Elicitation
Explanation schemes in MCDA context
Pairwise comparisons
Ordinal Sorting
Future prospects and applications
12/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Our context : Multiple Criteria Decision Aiding
DecisionMaker
A performance table, describing several actions
according to various criteria - the higher the better
A decision problem : Is action A better
than action B? Is action C good enough?
Sparse preferences between some actions
13/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications
Pairwise comparisons (choice or ranking)
DecisionMaker
I want to compare hotels described by 4 criteria :
A - comfort : (4?) A � a (2
?)
B - restaurant : (presence) B � b (absence)
C - commute time : (15 min) C � c (45 min)
D - cost : (50 $) D � d (150 $)
I prefer [AbCd] to [aBcD], [abcD]
to [aBCd] and [aBCd] to [Abcd]
I want to know :
Is [abCD] better than [ABcd] ?
14/44
Motivations Introduction to Multiple Criteria Decision Aiding Explanation schemes in MCDA context Future prospects and applications