Program

The program can be downloaded by clicking here

Keynote speakers:

• Denis Bouyssou, Université Paris Dauphine (Slides)
  Title: Preference modelling for multiattributed alternatives: the conjoint measurement approach
  Asbtract: The purpose of this talk is to offer a brief and nontechnical introduction to the main preference models for alternatives evaluated on several attributes that have been developed in the field of conjoint measurement. The emphasis is on the additive value function model.  We outline its axiomatic foundations  and present various possible assessment techniques to implement it.  Some extensions of this model, e.g., nonadditive models or models tolerating intransitive preferences will then be briefly reviewed.
• Nicolas Usunier, Facebook AI research (Slides)
  Title: Recommender systems and fair allocation
  Abstract: Recommender systems decide who sees what. They implement allocation mechanisms, where the scarce resource is the user’s attention, which many agents compete for: friends, creators, newspapers, advertisers, recruiters, and of course users themselves, who want valuable content in return for the time they spend on the platform. The literature on multi-stakeholder recommender systems emphasizes the need for fairness in this allocation process, in particular with respect to creators – avoiding disparities in exposure of creator groups that do not have a clear justification, or preventing winner-take-all effects of ranking pipelines. Promoting small creators has also been proposed to incentivize the creation of new content, thereby sustaining the ecosystem in the long run.
  The litterature on fairness in recommender systems (or fairness of exposure) routinely mentions the link with fair allocation mechanisms studied in economics, in particular in social choice theory and more specifically fair division. However, the algorithms that have been proposed so far mostly focus on (hard or soft) fairness constraints’’ such as equalizing exposure to users across various creator groups, which fail to satisfy basic fairness properties studied in fair division. In this talk I will describe our work to bring fairness criteria studied in social choice and fair division to recommender systems. In addition to alternative approaches to implement fairness towards creators and/or viewers, we propose new, computationally efficient algorithms for fair online ranking and exploration/exploitation.
• Thomas Augustin, Ludwig Maximilian University of Munich (Slides)
  Title: Some insights from decision making under strict uncertainty for machine learning
  Abstract: This presentation discusses whether transferring recent developments in decision making to machine learning may provide new opportunities and insights. In decision making under strict uncertainty, appropriately reflecting the underlying weakly structured information needs principled generalizations of the basic decision-theoretic concepts. Consequently, one works with partial orders, imprecise prior probabilities, sets of cardinal utilities, and optimal solutions consisting of a non-singleton set of actions.
  The first part of the presentation consists of an informal tour revisiting several issues and variants of machine learning methods where set-based concepts could be beneficial. In the second part, I will discuss one concrete, recently developed framework yielding a generalized notion of stochastic dominance. This notion allows ranking classifiers under multiple quality criteria and benchmark data sets while powerfully circumventing the cumbersome and possibly self-contradictory reliance on aggregates.