• 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.

    8:30–9:10REGISTRATION and wel­come (Inno­va­tion center)
    9:10–9:20Con­fer­ence opening
    9:20–10:50Pref­er­ence learn­ing (1)
    Chair: Thomas Augustin

    Improv­ing pref­er­ence learn­ing for MR-Sort using GPU (Paper)

    A Dual Approach for Learn­ing Sparse Rep­re­sen­ta­tions of Cho­quet Inte­grals (Paper)

    Find­ing Opti­mal Arms in Non-sto­chas­tic Com­bi­na­to­r­i­al Ban­dits with Appli­ca­tions in Algo­rithm Con­fig­u­ra­tion (Paper)
    10:50–11:10COFFEE BREAK
    11:10–12:10Thomas Augustin Keynote: “Some insights from deci­sion mak­ing under strict uncer­tain­ty for machine learn­ing” (Slides)
    Chair: Sébastien Destercke
    12:10–13:30LUNCH BREAK (Show­room)
    13:30–14:30Nico­las Usunier Keynote: “Rec­om­mender sys­tems and fair allo­ca­tion” (Slides)
    Chair: Ben­jamin Quost
    14:30–15:30Pref­er­ence learn­ing (2)
    Chair: Nico­las Usunier
    A genet­ic algo­rithm for learn­ing the para­me­ters of an SRMP pref­er­ence mod­el: Bastien Pas­de­loup, Arwa Khan­nous­si, Alexan­dru-Liviu Olteanu and Patrick Mey­er (Paper)

    Uni­ver­sal aggre­ga­tion of per­mu­ta­tions: Ekhine Iruroz­ki and Stephan Clemen­con (Paper)
    15:30–15:50COFFEE BREAK
    15:50–17:20Order­ing alter­na­tives
    Chair: Patrick Mey­er
    Exper­i­men­tal eval­u­a­tion of sim­i­lar­i­ty index on the MIA mol­e­c­u­lar fam­i­ly: Mejri Mohamed Yas­sine, Ozturk Meltem, Olivi­er Cail­loux and Benid­dir Ahmed Meh­di (Paper)

    Mul­ti­ple cri­te­ria sort­ing and max­i­mal antichains: Denis Bouys­sou, Thier­ry Marchant and Marc Pir­lot (Paper)

    A Char­ac­ter­i­za­tion of Choice Func­tions Rep­re­sentable by Pare­to-Embed­ding of Alter­na­tives: Karl­son Pfannschmidt and Eyke Hüller­meier (Paper)
    18:45Short car trav­el to Gala Din­ner (stop at train sta­tion and at Guy Denielou bus stop), then gala din­ner
    9:30–11:00Denis Bouys­sou Keynote: “Pref­er­ence mod­el­ling for mul­ti­at­trib­uted alter­na­tives: the con­joint mea­sure­ment approach” (Slides)
    Chair: Marc Pirlot
    11:00–11:20COFFEE BREAK
    11:20–12:20Space and time in pref­er­ences
    Chair: Khaled Belahcene
    Pref­er­ence dis­ag­gre­ga­tion involv­ing time-series: Beta­nia Campel­lo, Sarah Ben Amor, Leonar­do Tomazeli Duarte and João Romano (Paper)

    Learn­ing the para­me­ters of a mul­ti­di­men­sion­al spa­tial pref­er­ence mod­el in mul­ti-cri­te­ria deci­sion aid­ing: Arwa Khan­nous­si, Antoine Rol­land and Julien Vel­cin (Paper)
    13:30–15:00Explain­ing pref­er­ences
    Chair: Vin­cent Mousseau
    On the Expla­na­tion of a Gen­er­al Deci­sion Mod­el: Christophe Labreuche, Nico­las Atien­za and Roman Bres­son (Paper)

    Frag­ment of the nec­es­sary deci­sions explained by an argu­ment scheme in a non com­pen­sato­ry sort­ing mod­el: Khaled Belahcene, Jérome Gaigne and Syl­vain Lagrue (Paper)

    Explain­ing Fair­ness-Ori­ent­ed Rec­om­men­da­tions using Trans­fers and Tran­si­tive Argu­ments: Hénoïk Willot, Khaled Belahcene and Sébastien Dester­cke (Paper)
    15:00–16:00Good­byes and gen­er­al discussions