Le 12/06/2018 par Webmaster ROADEF :

From the first expert systems to the recent recommender systems which flourish on commercial

websites or are available in the context business-to-business sales applications, decision-
aiding has been a central concern in Artificial Intelligence. It has become clear that providing

recommendations is only part of the challenge: as decision-aiding tools become everyday more
popular and more sophisticated, it is of utmost importance to develop their explanatory
capabilities. Consequently, “Why should I trust you?” is a question every system should be

prepared to answer - especially when the stakes are high. Moreover, in the business-to-
business sales scenario, the explanation of how a given recommendation has been reached is

often more important than the recommendation itself: it gives the seller appropriate arguments
to approach the client. Finally, there is a growing demand of institutions and citizens to make
algorithmic decisions transparent and trustworthy. Indeed, the recent General Data Protection
Regulation (GDPR) adopted by the European Parliament goes further by adding a “right to
explanation”.
As a recommender system is targeted to an end-user, it must take into account the user’s
preferences and specificities. Therefore, such preferences have first to be elicited and then
used to constrain the recommendations. In addition, they can also be used to better formulate
and reinforce the explanation.
Under such perspectives, our project addresses the problems of recommendations, where the
aim for an “artificial agent adviser” is to help a human user (a decision maker) in building and
understanding the recommendations of a particular decision problem. Decision aiding is thus a
situation involving two parties: a user, with preferences which may be very incompletely defined
or difficult to convey, and an agent, which will have the capabilities of representing explicitly
and accountably the reasons for which it recommends a solution to a user. We are interested
by decision problems involving preference information and data carefully designed and
elaborated on human data sets. However, such data (e.g. preferences, values, etc.) could be
incomplete, imprecise, issued from different potentially conflicting sources. The
recommendations are based on Multiple Criteria Decision Aiding models that are well founded
from the point of view of Decision Theory [1].
An interesting starting point will be to consider the Non-Compensatory model (NCS) [2] and the
Simplified Ranking with Multiple Points (S-RMP) [3]. In short, the first model allows to classify
a set of options into an ordered predefined categories, while the second one has the aim to
construct a ranking among a set of options. These two models have the advantage to be

interpretable in the sense that the decision rules are simple and easy to grasp by a decision
maker [4,5].
We will investigate several use cases using both public and real-world industrial data sets. Also
we will demonstrate how explanations improve optimization models generated by Cognitive 
Optimization.
The Ph.D. student will be hosted by IBM France Lab in Saclay and the thesis is run in
collaboration between CentraleSupélec in the context of DATAIA Convergence Institute, IBM
France Lab in Saclay and IBM Research Zurich.
Candidates should have a Master Degree in Computer Science from an University or “grande
école”.He/she should have a good background in artificial intelligence and operational research.
Knowledge on multiple criteria decision analysis will be appreciated. Candidates are invited to
send their CV, motivation letter, their grades obtained in Master/Grande Ecole, and references.
Application deadline : 30/06/2018
Contacts :
• IBM France Lab : Christian de Sainte Marie, csma@fr.ibm.com
• IBM Research : Michalis Vlachos, Paolo Scotton {mvl,psc}@zurich.ibm.com
• CentraleSupélec : Wassila Ouerdane, Vincent Mousseau,
{wassila.ouerdane,vincent.mousseau}@centralesupelec.fr

References
[1] D. Bouyssou, T. Marchant, M. Pirlot, A. Tsoukiàs and P. Vincke, Evaluation and Decision
models: stepping stones for the analyst, Berlin, Springer Verlag, 2006.
[2] D. Bouyssou and T. Marchant, An axiomatic approach to Non compensatoty sorting
methods in mcdm I : the case of two categories, European Journal of Operation Research,
2007, 178(1), 217-245.
[3] A. Rolland, Reference-Based preference aggregation procedure in multi-criteria decision
making, European Journal of Operation Research, 2013, 225(3), 479-486.
[4] V. Ferreti, J. Liu, V. Mousseau and W. Ouerdane, Reference-based ranking procedure for
environmental decision making: Insights from an ex-post analysis, Environmental Modelling &
Software, 2018, 99, 11-24.
[5] O. Sobrie, M. Lazouni, S. Mahmoudi, V. Mousseau and M. Pirlot, A new decision support
model for preanesthetic evaluation, Computer Methods and Programs in Biomedicine, 2016,
133, 183-196.