2023Activity reportProjectTeamFAIRPLAY
RNSR: 202224251U Research center Inria Saclay Centre at Institut Polytechnique de Paris
 In partnership with:Institut Polytechnique de Paris, Criteo
 Team name: Coopetitive AI: Fairness, Privacy, Incentives
 In collaboration with:Centre de Recherche en Economie et Stastistique
 Domain:Applied Mathematics, Computation and Simulation
 Theme:Optimization, machine learning and statistical methods
Keywords
Computer Science and Digital Science
 A4.8. Privacyenhancing technologies
 A8.11. Game Theory
 A9.2. Machine learning
 A9.9. Distributed AI, Multiagent
Other Research Topics and Application Domains
 B9.9. Ethics
 B9.10. Privacy
1 Team members, visitors, external collaborators
Research Scientists
 Patrick Loiseau [Team leader, INRIA, Researcher, HDR]
 Marc Abeille [Criteo, Industrial member]
 Benjamin Heymann [Criteo, Industrial member]
 Hugo Richard [Criteo, Industrial member]
 Maxime Vono [Criteo, Industrial member]
Faculty Members
 Vianney Perchet [Team leader, Criteo & ENSAE , Professor, HDR]
 Cristina Butucea [ENSAE, Professor, HDR]
 Matthieu Lerasle [ENSAE, Professor, HDR]
PostDoctoral Fellows
 Dorian Baudry [ENSAE]
 Simon Finster [INRIA, PostDoctoral Fellow, from Oct 2023]
 Simon Finster [CNRS, PostDoctoral Fellow, until Sep 2023]
 Solenne Gaucher [ENSAE, from Sep 2023]
 Nadav Merlis [ENSAE]
 Denis Sokolov [INRIA, PostDoctoral Fellow, from Dec 2023]
PhD Students
 Ziyad Benomar [ENSAE]
 Maria Cherifa [Criteo, CIFRE]
 Lorenzo Croissant [ENSAE, ATER]
 Hafedh El Ferchichi [ENSAE]
 Côme Fiegel [ENSAE]
 Mike Liu [ENSAE]
 Mathieu Molina [INRIA]
 Matilde Tullii [ENSAE]
Technical Staff
 Felipe Garrido Lucero [INRIA, Engineer]
 Sruthi Gorantla [INRIA, Engineer, from Jun 2023 until Aug 2023]
Interns and Apprentices
 Reda Jalal [INRIA, Intern, from May 2023 until Nov 2023]
 Giovanni Montanari [ENSAE, Intern, from Apr 2023 until Sep 2023]
 Nicolas Noldus [INRIA, Intern, from Jun 2023 until Jul 2023]
Administrative Assistant
 Melanie Da Silva [Inria, from May 2023]
External Collaborators
 Clément Calauzènes [Criteo]
 Remi Castera [Univ. Grenoble Alpes]
 Julien Combe [Ecole Polytechnique, HDR]
 Azadeh Khaleghi [ENSAE]
 Jaouad Mourtada [ENSAE]
2 Overall objectives
2.1 Broad context
One of the principal objectives of Machine Learning (ML) is to automatically discover using past data some underlying structure behind a data generating process in order either to explain past observations or, perhaps more importantly, to make predictions and/or to optimize decisions made on future instances. The area of ML has exploded over the past decade and has had a tremendous impact in many application domains such as computer vision or bioinformatics.
Most of the current ML literature focuses on the case of a single agent (an algorithm) trying to complete some learning task based on gathered data that follows an exogenous distribution independent of the algorithm. One of the key assumptions is that this data has sufficient “regularity” for classical techniques to work. This classical paradigm of “a single agent learning on nice data”, however, is no longer adequate for many practical and crucial tasks that imply users (who own the gathered data) and/or other (learning) agents that are also trying to optimize their own objectives simultaneously, in a competitive or conflicting way. This is the case, for instance, in most learning tasks related to Internet applications (content recommendation/ranking, ad auctions, fraud detection, etc.). Moreover, as such learning tasks rely on users' personal data and as their outcome affect users in return, it is no longer sufficient to focus on optimizing prediction performance metrics—it becomes crucial to consider societal and ethical aspects such as fairness or privacy.
The field of single agent ML builds on techniques from domains such as statistics, optimization, or functional analysis. When different agents are involved, a strategic aspect inherent in game theory enters the picture. Indeed, interactions—either positive or negative—between rational entities (firms, single user at home, algorithms, etc.) foster individual strategic behavior such as hiding information, misleading other agents, freeriding, etc. Unfortunately, this selfishness degrades the quality of the data or of the predictions, prevents efficient learning and overall may diminish the social welfare. These strategic aspects, together with the decentralized nature of decision making in a multiagent environment, also make it harder to build algorithms that meet fairness and privacy constraints.
The overarching objective of FAIRPLAY is to create algorithms that learn for and with users—and techniques to analyze them—, that is to create procedures able to perform classical learning tasks (prediction, decision, explanation) when the data is generated or provided by strategic agents, possibly in the presence of other competing learning agents, while respecting the fairness and privacy of the involved users. To that end, we will naturally rely on multiagent models where the different agents may be either agents generating or providing data, or agents learning in a way that interacts with other agents; and we will put a special focus on societal and ethical aspects, in particular fairness and privacy. Note that in FAIRPLAY, we focus on the technical challenges inherent to formalizing mathematically and respecting ethical properties such as nondiscrimination or privacy, often seen as constraints in the learning procedure. Nevertheless, throughout the team's life, we will reflect on these mathematical definitions for the particular applications studied, in particular their philosophical roots and legal interpretation, through interactions with HSS researchers and with legal specialists (from Criteo).
2.1.1 Multiagent systems
Any company developing and implementing ML algorithms is in fact one agent within a large network of users and other firms. Assuming that the data is i.i.d. and can be treated irrespectively of the environment response—as is done in the classical ML paradigm—might be a good first approximation, but should be overcome. Users, clients, suppliers, and competitors are adaptive and change their behavior depending on each other's interactions. The future of many ML companies—such as Criteo—will consist in creating platforms matching the demand (created by their users) to the offer (proposed by their clients), under the system constraints (imposed by suppliers and competitors). Each of these agents have different, conflicting interests that should be taken into account in the model, which naturally becomes a multiagent model.
Each agent in a multiagent system may be modeled as having their own utility function ${u}_{i}$ that can depend on the action of other agents. Then, there are two main types of objectives: individual or collective 99. If each agent is making their own decision, then they can be modeled as each optimizing their own individual utility (which may include personal benefit as well as other considerations such as altruism where appropriate) unilaterally and in a decentralized way. This is why a mechanism providing correct incentives to agents is often necessary. At the other extreme, social welfare is the collective objective defined as the cumulative sum of utilities of all agents. To optimize it, it is almost always necessary to consider a centralized optimization or learning protocol. A key question in multiagent systems is to apprehend the “social cost” of letting agents optimize their own utility by choosing unilaterally their decision compared to the one maximizing social welfare; this is often measured by the “price of anarchy”/“price of stability” 108: the ratio of the maximum social welfare to the (worst/best) social welfare when agents optimize individually.
The natural language to model and study multiagent systems is game theory—see below for a list of tools and techniques on which FAIRPLAY relies, game theory being the first of them. Multiagent systems have been studied in the past; but not with a focus on learning systems where agents are either learning or providing data, which is our focus in FAIRPLAY and leads to a blend of game theory and learning techniques. We note here again that, wherever appropriate, we shall reflect (in part together with colleagues from HSS) on the soundness of the utility framework for the considered applications.
2.1.2 Societal aspects and ethics
There are several important ethical aspects that must be investigated in multiagent systems involving users either as data providers or as individuals affected by the ML agent decision (or both).
Fairness and Discrimination
When ML decisions directly affect humans, it is important to ensure that they do not violate fairness principles, be they based on ethical or legal grounds. As ML made its way in many areas of decision making, it was unfortunately repeatedly observed that it can lead to discrimination (regardless of whether or not it is intentional) based on gender, race, age, or other sensitive attributes. This was observed in online targeted advertisement 93, 119, 33, 77, 93, 35, but also in many other applications such as hiring 64, datadriven healthcare 71, or justice 94. Biases also have the unfortunate tendency to reinforce. An operating multiagent learning system should be able in the long run to get rid by itself of inherent population biases, that is, be fair amongst users irrespective of the improperly constructed dataset.
The mathematical formulation of fairness has been debated in recent works. Although a few initial works proposed a notion of individual fairness, which mandates that “similar individuals” receive “similar outcomes” 66, this notion was quickly found unpractical because it relies on a metric to define closeness that makes the definition somewhat arbitrary. Most of the works then focused on notions of group fairness, which mandate equality of outcome “on average” across different groups defined by sensitive attributes (e.g., race, gender, religious belief, etc.). Most of the works on group fairness focus on the classification problem (e.g., classifying whether a job applicant is good or not for the job) where each data example $({X}_{i},{Y}_{i})$ contains a set of features ${X}_{i}$ and a true label ${Y}_{i}\in \{0,1\}$ and the goal is to make a prediction ${\widehat{Y}}_{i}$ based on the features ${X}_{i}$ that has a high probability to be equal to the true label. Assuming that there is a single sensitive attribute ${s}_{i}$ that can take two values $a$ or $b$, this defines two groups: those for whom ${s}_{i}=a$ and those for whom ${s}_{i}=b$. There are several different concepts of group fairness that can be considered; we shall especially focus on demographic parity (DP), which prescribes $P({\widehat{Y}}_{i}=1{s}_{i}=a)=P({\widehat{Y}}_{i}=1{s}_{i}=b)$ and equal opportunity (EO) 78, which mandates that $P({\widehat{Y}}_{i}=1{s}_{i}=a,{Y}_{i}=1)=P({\widehat{Y}}_{i}=1{s}_{i}=b,{Y}_{i}=1)$.
The fair classification literature proposed, for each of these fairness notions, ways to train fair classifiers based on three main ideas: preprocessing 127, inprocessing 125, 126, 122, and postprocessing 78. All of these works, however, focus on idealized situations where a single decisionmaker has access to ground truth data with the sensitive features and labels in order to train classifiers that respect fairness constraints. We use similar group fairness definitions and extend them (in particular through causality), but our goal is to go further in terms of algorithms by modeling practical scenarios with multiple decisionmakers and incomplete information (in particular lack of ground truth on the labels).
Privacy vs. Incentives
ML algorithms, in particular in Internet applications, often rely on users' personal information (whether it is directly their personal data or indirectly some hidden “type” – gender, ethnicity, behaviors, etc.). Nevertheless, users may be willing to provide their personal information if it increases their utility. This brings a number of key questions. First, how can we learn while protecting users' privacy (and how should privacy even be defined)? Second, finding the right balance between those two apriori incompatible concepts is challenging; how much (and even simply how) should an agent be compensated for providing useful and accurate data?
Differential privacy is the most widely used private learning framework 65, 67, 114 and ensures that the output of an algorithm does not significantly depend on a single element of the whole dataset. These privacy constraints are often too strong for economic applications (as illustrated before, it is sometimes optimal to disclose some private information). $f$divergence privacy costs have thus been proposed in recent literature as a promising alternative 57. These $f$divergences, such as KullbackLeibler, are also used by economists to measure the cost of information from a Bayesian perspective, as in the rational inattention literature 118, 101, 96. It was only recently that this approach was considered to measure “privacy losses” in economic mechanisms 68. In this model, the mechanism designer has some prior belief on the unobserved and private information. After observing the player's action, this belief is updated and the cost of information corresponds to the KL between the prior and posterior distributions of this private information.
This privacy concept can be refined up to a single user level, into the socalled local differential privacy. Informally speaking, the algorithm output can also depend on a single user data that still must be kept private. Estimation are actually sometimes more challenging under this constraint, i.e., estimation rates degrade 115, 52, 53 but is sometimes more adapted to handle usergenerated data 73.
Interestingly, we note that the notions of privacy and fairness are somewhat incompatible. This will motivate Theme 2 developed in our research program.
2.2 A large variety of tools and techniques
Analyzing multiagent learning systems with ethical constraints will require us to use, develop, and merge several different theoretical tools and techniques. We describe the main ones here. Note that although FAIRPLAY is motivated by practical usecases and applications, part of the team's objectives is to improve those tools as necessary to tackle the problems studied.
Game theory and economics
Game theory 72 is the natural mathematical tool to model multiple interacting decisionmakers (called players). A game is defined by a set of players, a set of possible actions for each player, and a payoff function for each player that can depend on the actions of all the players (that is the distinguishing feature of a game compared to an optimization problem). The most standard solution concept is the socalled Nash equilibrium, which is defined as a strategy profile (i.e., a collection of possibly randomized action for each player) such that each player is at best response (i.e., has the maximum payoff given the others' strategies). It is a “static” (oneshot) solution concept, but there also exist dynamic solution concepts for repeated games 56, 103.
Online and reinforcement learning 49
In online learning (a.k.a. multiarmed bandit 50, 110), data is gathered and treated on the fly. For instance, consider an online binary classification problem. Some unlabelled data ${X}_{t}\in {R}^{d}$ is observed, and the agent predicts its label ${Y}_{t}$; let us denote ${\widehat{Y}}_{t}\in \pm 1$ the prediction. The agent potentially observes the loss $1\{{Y}_{t}\ne {\widehat{Y}}_{t}\}$ and then receives another new unlabeled data example ${X}_{t+1}$. In that specific problem, the typical learning objective is to perform asymptotically as good as the best classifier ${f}^{*}$ in some given class $\mathcal{F}$, i.e., such that the loss ${\sum}_{t=1}^{T}1\{{Y}_{t}\ne {\widehat{Y}}_{t}\}$ is $o\left(T\right)$close to ${max}_{f\in \mathcal{F}}{\sum}_{t=1}^{T}1\{{Y}_{t}\ne f\left({X}_{t}\right)\}$; the difference between those terms is called regret. The more general model with an underlying state of the world ${S}_{t}\in \mathcal{S}$ that evolves at each step following some Markov Decision Process (MDP, i.e., the transition matrix from ${S}_{t}$ to ${S}_{t+1}$ depend on the actions of the agent) and impacts the loss function is called reinforcement learning (RL). RL is an incredibly powerful learning technique, provided enough data are available since learning is usually quite slow. This is why the recent successes involve settings with heavy simulations (like games) or wellunderstood physical systems (like robots).
These techniques will be central to our approach as we aim to model problems where ground truth data is not available upfront and problems involving sequential decision making. There have been some successful first results in that direction. For instance, there are applications (e.g., cognitive radio) where several agents (users) aim at finding a matching with resources (the different bandwidth). They can do that by “probing” the resources, estimating their preferences and trying to find some stable matchings 47, 95.
Online algorithms 45 and theoretical computer science
Online algorithms are closely related to online learning with a major twist. In online learning, the agent has “0look ahead”; for instance, in the online binary classification example, the loss at stage $t$ was $1\{{Y}_{t}\ne {\widehat{Y}}_{t}\}$ but ${Y}_{t}$ was not known in advance. The comparison class, on the other hand, was the empirical performance of a given set of classifiers. In online algorithms, the agents have “1look ahead”; in the classification example, this means that ${Y}_{t}$ is known before choosing ${\widehat{Y}}_{t}$. But the overall objective is obviously no longer the minimisation of the empirical error, but the minimisation of this error plus the total number of changes (say). The comparison class is then larger, namely a subset of admissible (or the whole set) sequences of prediction ${\{\pm 1\}}^{T}$. The typical and relevant example of online problem relevant for Criteo that will be investigated is the matching problem: agents and resources arrive sequentially and must be, if possible, paired together as fast as possible (and as successfully as possible). Variants of these problems include the optimal stopping time question (when/how make a final decision) such as prophet inequalities and related questions 62,
Optimal transport 120
Optimal transport is a quite old problem introduced by Monge where an agent aims at moving a pile of sand to fill a hole at the smallest possible price. Formally speaking, given two probability measures $\mu $ and $\nu $ on some space $\mathcal{X}$, the optimal transport problem consist in finding (if it exists, otherwise the problem can be relaxed) a transport map $T:\mathcal{X}\to \mathcal{X}$ that minimizes ${\int}_{\mathcal{X}}c(x,T\left(x\right))d\mu \left(x\right)$ for some cost function $c:{\mathcal{X}}^{2}\to R$, under the constraint that $T\u266f\mu =\nu $, where $T\u266f\mu $ is the pushforward measure of $\mu $ by $T$. Interestingly, when $\mu $ and $\nu $ are empirical measures, i.e., $\mu =\frac{1}{N}{\sum}_{n=1}{\delta}_{{x}_{n}}$ and $\nu =\frac{1}{N}{\sum}_{n=1}{\delta}_{{y}_{n}}$, a transport map is nothing more than a matching between $\left\{{x}_{n}\right\}$ and $\left\{{y}_{n}\right\}$ that minimizes the cost ${\sum}_{n}c({x}_{n},T\left({x}_{n}\right))$.
Recently, optimal transport gained a lot of interest in the ML community 109 thanks to its application to images and to new techniques to compute approximate matchings in a tractable way 112. Even more unexpected applications of optimal transport have been discovered: to protect privacy 48, fairness 41, etc. Those connections are promising, but only primitive for the moment. For instance, consider the problem of matching students to schools. The unfairness level of a school can be measured as the Wasserstein distance between the distribution of the students within that school compared to the overall distribution of students. Then the matching algorithms could have a constraint of minimizing the sum of (or its maximum) unfairness levels; alternatively, we could aim at designing mechanisms giving incentives to schools to be fair in their allocation (or at least in their list preferences), typically by paying a higher fee if the unfairness level is high.
2.3 General objectives
The overarching objective of FAIRPLAY of to create algorithms to learn for and with users—and techniques to analyze them—, through the study of multiagent learning systems where the agents can be cooperatively or competitively learning agents, or agents providing or generating data, while guaranteeing that fairness and privacy constraints are satisfied for the involved users. We detail this global objective into a number of more specific ones.

Objective 1: Developing fair and private mechanisms
Our first objective is to incorporate ethical aspects of fairness and privacy in mechanisms used in typical problems occurring in Internet applications, in particular auctions, matching, and recommendation. We will focus on social welfare and consider realistic cases with multiple agents and sequential learning that occur in practice due to sequential decision making. Our objective is both to construct models to analyze the problem, to devise algorithms that respect the constraints at stake, and to evaluate the different tradeoffs in standard notions of utility introduced by ethical constraints.

Objective 2: Developing multiagent statistics and learning
Data is now acquired, treated and/or generated by a whole network of agents interacting with the environment. There are also often multiple agents learning either collaboratively or competitively. Our second objective is to build a new set of tools to perform statistics and learning tasks in such environments. To this end, we aim at modeling these situations as multiagent systems and at studying the dynamics and equilibrium of these complex gametheoretic situations between multiple learning algorithms and data providers.

Objective 3: Improving the theoretical state of the art
Research must rely on theoretical, proven guarantees. We develop new results for the techniques introduced before, such as prophet inequalities, (online) matchings, bandits and RL, etc.

Objective 4: Proposing practical solutions and enhancing transfer from research to industry
Our last scientific objective is to apply and implement theoretical works and results to practical cases. This will be a crucial component of the project as we focus on transfer within Criteo.

Objective 5: Scientific Publications
We aim at publishing our results in toptier machine learning conferences (NeurIPS, ICML, COLT, ICLR, etc.) and in toptier game theory journals (Games and Economic Behavior, Mathematics of OR, etc.). We will also target conferences at the junction of those fields (EC, WINE, WebConf, etc.) as well as conferences specifically on security and privacy (IEEE S&P, NDSS, CSS, PETS, etc.) and on fairness (FAccT, AIES).
All the five objectives are interlaced. For instance, fairness and privacy constraints are important in Objective 2 whereas the multiagent aspect is also important in Objective 1. Objectives 4 and 5 are transversal and present in all the first three objectives.
3 Research program
To reach the objectives laid out above, we organize the research in three themes. The first one focuses on developing fair mechanisms. The second one considers private mechanisms, and in particular considers the challenge of reconciling fairness and privacy—which are often conflicting notions. The last theme, somewhat transverse to the first two, consists in leveraging/incorporating structure in all those problems in order to speed up learning. Of course, all themes share common points on both the problems/applications considered and the methods and tools used to tackle them; hence there are crossfertilization between the different themes.
3.1 Theme 1: Developing fair mechanisms for auctions and matching problems
3.1.1 Fairness in auctionbased systems
Online ads platforms are nowadays used to advertise not just products, but also opportunities such as jobs, houses, or financial services. This makes it crucial for such platforms to respect fairness criteria (be it only for legal reasons), as an unfair ad system would deprive a part of the population of some potentially interesting opportunities. Despite this pressing need, there is currently no technical solution in place to provably prevent discriminations. One of the main challenge is that ad impression decisions are the outcome of an auction mechanism that involves bidding decisions of multiple selfinterested agents controlling only a small part of the process, while group fairness notions are defined on the outcome of a large number of impressions. We propose to investigate two mechanisms to guarantee fairness in such a complex auctionbased system (note that we focus on online ad auctions but the work has broader applicability).

Advertisercentric (or biddercentric) fairness
We first focus on advertisercentric fairness, i.e., the advertiser of a thirdparty needs to make sure that the reached audience is fair independently of the ad auction platform. A key difficulty is that the advertiser does not control the final decision for each ad impression, which depends on the bids of other advertisers competing in the same auction and on the platform's mechanism. Hence, it is necessary that the advertiser keeps track of the auctions won for each of the groups and dynamically adjusts its bids in order to maintain the required balance.
A first difficulty is to model the behavior of other advertisers. We can first use a meanfield games approach similar to 81 that approximates the other bidders by an (unknown) distribution and checks equilibrium consistency; this makes sense if there are many bidders. We can also leverage refined meanfield approximations 75 to provide better approximations for smaller numbers of advertisers. Then a second difficulty is to find an optimal bidding policy that enforces the fairness constraint. We can investigate two approaches. One is based on an MDP (Markov Decision Process) that encodes the current fairness level and imposes a hard constraint. The second is based on modeling the problem as a contextual bandit problem. We note that in addition to fairness constraints, privacy constraints may complicate the optimal solution finding.

Platformcentric (or auctioncentric) fairness
We also consider the problem from the platform's perspective, i.e., we assume that it is the platform's responsibility to enforce the fairness constraint. We also focus here on demographic parity. To make the solution practical, we do not consider modification of the auction mechanism, instead we consider a given mechanism and let the platform adapt dynamically the bids of each advertiser to achieve the fairness guarantee. This approach would be similar to the pacing multipliers used by some platforms 61, 60, but using different multipliers for the different groups (i.e., different values of the sensitive attribute).
Following recent theoretical work on auction fairness 54, 80, 58 (which assumes that the targeted population of all ads is known in advance along with all their characteristics), we can formulate fairness as a constraint in an optimization problem for each advertiser. We study fairness in this static auction problem in which the auction mechanism is fixed (e.g., to second price). We then move to the online setting in which users (but also advertisers) are dynamic and in which decisions must be taken online, which we approach through dynamic adjustment of pacing multipliers.
3.1.2 Fairness in matching and selection problems
In this second part, we study fairness in selection and matching problems such as hiring or college admission. The selection problem corresponds to any situation in which one needs to select (i.e., assign a binary label to) data examples or individuals but with a constraint on the number of maximum number of positive labels. There are many applications of selection problems such as police checks, loan approvals, or medical screening. The matching problem can be seen as the more general variant with multiple selectors. Again, a particular focus is put here on cases involving repeated selection/matching problems and multiple decision makers.

Fair repeated multistage selection
In our work 69, we identified that a key source of discrimination in (static) selection problems is differential variance, i.e., the fact that one has quality estimates that have different variances for different groups. In practice, however, the selection problem is often ran repeatedly (e.g., at each hiring campaign) and with partial (and increasing) information to exploit for making decisions.
Here, we consider the repeated multistage selection problem, where at each round a multistage selection problem is solved. A key aspect is that, at the end of a round, one learns extra information about the candidates that were selected—hence one can refine (i.e., decrease the variance of) the quality estimate for the groups in which more candidates were selected. We will first rethink fairness constraints in this type of repeated decision making problems. Then we will both study the discrimination that come out of natural (e.g., greedy) procedure as well as design (near) optimal ones for the constraints at stake. We also investigate how the constraints affect the selection utility.

Multiple decisionmakers
Next, we investigate cases with multiple decisionmakers. We propose two cases in particular. The first one is the simple twostage selection problem but where the decisionmaker doing the firststage selection is different from the decisionmaker doing the secondstage selection. This is a typical case for instance for recruiting agencies that propose sublists of candidates to different firms that wish to hire. The second case is when multipledecision makers are trying to make a selection simultaneously—a typical example of this being the college admission problem (or faculty recruitment). We intend to model it as a game between the different colleges and to study both static solutions as well as dynamic solutions with sequential learning, again modeling it as a bandit problem and looking for regretminimizing algorithms under fairness constraints. A number of important questions arise here: if each college makes its selection independently and strategically (but based on quality estimates with variances that differ amongst groups), how does it affect the “global fairness” metrics (meaning in aggregate across the different colleges) and the “local fairness” metrics (meaning for an individual college)? What changes if there is a central admission system (such as Parcoursup)? And in this later case, how to handle fairness on the side of colleges (i.e., treat each college fairly in some sense)?

Fair matching with incentives in twosided platforms
We will study specifically the case of a platform matching demand on one side to offer on the other side, with fairness constraints on each side. This is the case for instance in online job markets (or crowdsourcing). This is similar to the previous case but, in addition, here there is an extra incentives problem: companies need to give the right incentives to job applicants to accept the jobs; while the platform doing the match needs to ensure fairness on both sides (job applicants and companies). This gives rise to a complicated interplay between learning and incentives that we will tackle again in the repeated setting.
We finally mention that, in many of these matching problems, there is an important time component: each agent needs to be matched “as soon as possible”, yielding a tradeoff between the delay to be matched and the quality of the match. There is also a problem of participation incentives; that is how the matching algorithm used affect the behavior of the participants in the matching “market” (participation or not, information revelation, etc.). In the longterm, we will incorporate these aspects in the above models.
Throughout the work in this theme, we will also consider a question transverse and present in all the models above: how can we handle multidimensional fairness, that is, where there are multiple sensitive attributes and consequently an exponential number of subgroups defined by all intersections; this combinatorial is challenging and, for the moment, still exploratory.
3.2 Theme 2: Reconciling, and enforcing privacy with fairness
In the previous theme, we implicitly assumed that we know the users' group, i.e., their sensitive attributes such as gender, age, or ethnicity. In practice, one of the key question when implementing fairness mechanisms is how to measure/control fairness metrics without having access to these protected attributes. This question relates to the link between privacy and fairness and the tradeoff between them (as fairness requires data and privacy tends to protect it) 113, 63.
A first option to solve this problem would be (when it is possible) to build proxies 76, 116 for protected attributes using available information (e.g., websites visited or products bought) and to measure or control for fairness using those in place of the protected attributes. As the accuracy of these proxies cannot be assessed, however, they cannot be used for any type of “public certification”—that is, for a company to show reasonable fairness guarantees to clients (e.g., as a commercial argument), or (even less) to regulators. Moreover, in many cases, the entity responsible for fairness should not be accessing sensitive information, even through proxies, for privacy reasons.
In FAIRPLAY, we investigate a different means of certifying fairness of decisions without having access to sensitive attributes, by partnering with a trusted thirdparty that collects protected attributes (that could for instance be a regulator or a public entity, such as Nielsen, say). We distinguish two cases:
 If the thirdparty and the company share a common identifier of users, then computing the fairness metric without leaking information to each other will boil down to a problem of secure multiparty computation (SMC). In such a case, there could be a need to be able to learn, which opens the topic of learning and privacy under SMC. This scenario, however, is likely not the most realistic one as having a common identifier requires a deep technical integration.
 If the thirdparty and the company do not share a common identifier of users, but there are common features that they both observe 84, then it is possible only to partially identify the joint distribution. With additional structural assumptions on the distribution, however, it could be identified accurately enough to estimate biases and fairness metrics. This becomes a distribution identification problem and brings a number of questions such as: how to do the distribution identification? how to optimally query data from the third party to train fair algorithms with high utility? etc. An important point to keep in mind in such a study is that it is likely that the third party userbase is different from that of the company. It will therefore be key to handle the covariate shift from one distribution to the other while estimating biases.
This distribution identification problem will be important in the context of privacy, even independently of fairness issues. Indeed, in the near future, most learning will happen in a privacypreserving setting (for instance, because of the Chrome privacy sandbox). This will require new learning schemes (different from e.g., Empirical Risk Minimization) as samples from the usual joint distribution $(X,Y)$ of samples/labels will no longer be observed. Only aggregated data—e.g., (empirical) marginals of the form $E\left[Y\right{X}_{2}=4,{X}_{7}=\text{``lemonde.fr''}]$—will be observed, with a limited budget of requests. This also brings questions such as how to mix it with ERM on some parts of the traffic, what is the (certainly adaptive or active) optimal strategy to query the marginals, etc. This problem will be further complicated by the fact that privacy (for instance through the variety of consents) will be heterogeneous: all features are not available all the time. This is therefore strongly related to learning with missing features and imputation 83.
In relation to the above problems, a key question is to determine what is the most appropriate definition of fairness to consider. Recall that it is wellknown that usual fairness metrics are not compatible 88. Moreover, in online advertising, fairness can be measured at multiple different levels: at the level of bids, at the level of audience reached, at the level of clicking users, etc. Fairness at the level of bids does not imply fairness of the audience reached (see Theme 1); yet external auditors would measure which ad is displayed—as was done for some ad platforms 117—hence in terms of public image, that would be the appropriate level to define fairness. Intimately, the above problem relates to the question of measuring what is the relevant audience of an ad, which would define the label if one were to use the EO fairness notion. This label is typically not available. We can explore three ways to overcome this issue. The first is to find a sequential way to learn the label through users clicking on ads. The second and third options are to focus in a first step on individual fairness, or on counterfactual fairness 91, which has many possible different level of assumptions and was popularized in 2020 92. The notion of counterfactual is key in causality 111. A model is said counterfactually fair if its prediction does not change (too much) by intervening on the sensitive attribute. Several works already propose ways of designing models that are counterfactually fair 87, 124, 123. This seems to be quite an interesting, but challenging direction to follow.
Finally, an alternative direction would be to purse modeling the tradeoff between privacy and fairness. For instance, in some game theoretic models, users can choose the quantity of data that they reveal 74, 48, so that the objective functions integrate different levels of fairness/privacy. Then those models model should be studied both in terms of equilibrium and in the online setup, with the objective of identifying how the strategic privacy considerations affect the fairnessutility tradeoff.
3.3 Theme 3: Exploiting structure in online algorithms and learning problems
Our last research direction is somewhat transverse, with possible application to improving algorithms in the first three themes. We explore how the underlying structure can be exploited, in the online and learning problems considered before, to improve performance. Note that, in all these problems, we will incorporate the fairness and privacy aspects even if they are somewhat transverse to the structure considered.1 The following sections are illustrating examples on how hidden structure can be leveraged in specific examples.
3.3.1 Leveraging structure in online matching
Finding large matchings in graphs is a longstanding problem with a rich history and many practical and theoretical applications 102, 79, 39, 38. Recall that given a graph $G=\left(\mathcal{E}\right)$—where $isasetofverticesand$E$isasetofedges,amatching$ME$isasubsetofedgessuchthateachvertexbelongstoatmostoneedge$eM$.Inthatcontext,aperfectmatching$M$isamatchingwhereeachvertex$v is associated to an edge $e\in \mathcal{M}$, and a maximum matching is a matching of maximum size (one can also consider weights on edges). Here, we study an online setting, which is more adequate in applications such as Internet advertising where ad impressions must be assigned to available ad slots 102, 51. Consider a bipartite graph, where $U\cup V$ is the union of two disjoints sets. Nodes $u\in U$ are known beforehand, whereas nodes $v\in V$ are discovered one at a time, along with the edges they belong to, and must be either immediately matched to an available (i.e., unmatched yet) vertex $u\in U$ or discarded. Online bipartite matching is relevant in twosided markets besides ad allocations such as assigning tasks to workers 79.
A natural measure for the quality of an online matching algorithm is the “competitive ratio” (CR): the ratio between the size of the created matching to the size of the optimal one 102. The seminal work 86 introduced an optimal algorithm for the adversarial case 42, that guarantees a CR of $1\frac{1}{e}$; but focusing on a pessimistic worstcase. In practice, some relevant knowledge (either given a priori or learned during the process) on the underlying structure of the problem can be leveraged. The focus then shifted to models taking into account some type of stochasticity in the arrival model, mostly for the i.i.d. model where arriving vertices $v\in V$ are drawn from a fixed distribution $\mathcal{D}$82, 40, 70, 85, 97, 98. The classical approach consists in optimizing the CR over the distribution $\mathcal{D}$. Even in this seemingly optimistic framework, however, it is now known that there is no hope for a CR of more than 0.823 98. Moreover, this generally leads to very large linear programs (LP).
A more recent approach restricts the distribution $\mathcal{D}$ over which the problem is optimized to classes of graphs with an underlying stochastic structure. The benefit of this approach is twofold: it gives hope for higher competitive ratios, and for simpler algorithms. Experiments also proved that complex algorithms optimized on $\mathcal{D}$ fared no better than simple greedy heuristics on “reallife” graphs 46. A few results along these lines show that is a promising path. For instance, 51 studied the problem on graphs where each vertex has degree at least $d$ and found a competitive ratio of $1{(11/d)}^{d}$. On dregular graphs, 59 designed a $1O(\sqrt{logd}/\sqrt{d})$ competitive algorithm. 100 showed that greedy algorithms were highly efficient on ErdösRenyi random graphs, with a competitive ratio of 0.837 in the worst case. 37 showed that in a specific market with two types of matching agents, the behavior of the matching algorithm varies with the homogeneity of the market. Our goal here is to go beyond the independence assumption underlying all these works.

Introducing correlation and inhomogeneity
We will start by deriving and studying optimal online matching strategies on widely studied classes of graphs that present simple inhomongeneity or correlation structures (which are often present in applications). The stochastic block model 34 is often used to generate graphs with an underlying community structure. It presents a simple correlation structure: two vertices in the same community are more likely to have a common neighbors than two vertices in different communities. Another focus point will be a generalized version of the ErdösRenyi model, where the inplace vertices $u\in \mathcal{U}$ are divided into sets ${s}_{i}$, where $u\in {s}_{i}$ generates an edge with probability ${p}_{i}=\frac{{c}_{i}}{n}$. These two settings should give us a better understanding of how heterogeneity and correlation affect the matching performance.
Deriving the competitive ratio implies to study the asymptotic size of maximum matchings in random graphs. Two methods are usually used. The first and constructive one is the study of the KarpSipser algorithm on the graph 43. The second one involves the rich theory of graph weak local convergence 44. A straightforward application of the methods, however, requires the graph to have independence properties; adapting them to graphs with a correlation structure will require new ideas.

Configuration models and random geometric graphs
A configuration model is described as follows (in the bipartite case). Each vertex $u\in \mathcal{U}$ has a number of halfedges drawn for the same distribution ${\pi}_{\mathcal{U}}$ and each vertex $v\in hasanumberofhalfedgesdrawnfrom$V$(withtheassumptionthattheexpectedtotalnumbersofhalfedgesfrom$U$and$ are the same). Then a vertex $v\in thatarrivesinthesequentialfashionhasitshalfedges``complete{d}^{\text{'}\text{'}}bya\left(stillfree\right)halfedgeof$U$.Thisisastandardwayofcreatingrandomgraphswith\left(almost\right)fixeddistributionofdegrees.Herethequestionwouldsimplybethecompetitiveratioofsomegreedyalgorithm,whetherthedistributions$U$and$$areknownbeforehandorlearnedonthefly.Aninterestingvariantofthisproblemwouldbetoassumetheexistenceofa\left(hiddenornot\right)geometricgraph.Each$u U$isdrawni.i.d\phantom{\rule{4pt}{0ex}}in$Rd$\left(sayaGaussiancenteredat0\right)andsimilarlyfor$v . Then there is an edge between $u$ and $v$ with a probability depending on the distance between them. Here again, interesting variants can be explored depending on whether the distribution is known or not, and whether the locations of $u$ and/or $v$ are observed or not.

Learning while matching
In practical applications, the full stochastic structure of the graphs may not be known beforehand. This begs the question: what will happen to the performance of the algorithms if the graph parameters are learned while matching? In the generalized ErdösRenyi graph, this will correspond to learning the probability of generating edges. For the stochastic block model, the matching algorithm will have to perform online community detection.
3.3.2 Exploiting sideinformation in sequential learning
We end with an open direction that may be relevant to many of the problems considered above: how to use sideinformation to speedup learning. In many sequential learning problems where one receives feedback for each action taken, it is actually possible to deduce, for free, extra information from the known structure of the problem. However, how to incorporate that information in the learning process is often unclear. We describe it through two examples.

Onesided feedback in auctions
In online ad auctions, the advertisers' strategy is to bid in a compact set of possible bids. After placing a bid, the advertiser learns whether they won the auction or not; but even if they do not observe the bids of other advertisers, they can deduce for free some extra information: if they win they learn that they would have won with any higher bid and if they loose they learn that they would have lost with any lower bid 121, 55. We will investigate how to incorporate this extra information in RL procedures devised in Theme 1. One option is by leveraging the KaplanMeier estimator.

Sideinformation in dynamic resource allocation problems and matching
Generalizing the idea above, one can observe sideinformation in many other problems 36. Typically, in resource allocation problems (e.g., how to allocate a budget of ad impressions), one can leverage a monotony property: one would not have gained more by allocating less. Similarly, in matching with unknown weights, it is often possible upon doing a particular match to learn the weight of other potential pairs.
4 Application domains
4.1 Typical problems and usecases
In FAIRPLAY, we focus mainly on problems involving learning that relate to Internet applications. We will tackle generic problems (in particular auctions and matching) with different applications in mind, in particular some applications in the context of Criteo's business but also others. A crucial property of those applications is the aforementioned ethical concerns, namely fairness and/or privacy. The team was shaped and motivated by several such usecases, from more practical (with possible short or middle term applications in particular in Criteo products) to more theoretical and exploratory ones. We describe first here the main types of generic problems and usecases considered in this context.
Auctions 90
There are many different types of auctions that an agent can use to sell one or several items in her possession to $n$ potential buyers. This is the typical way in which spots to place ads are sold to potential advertisers. In case of a single item, the seller ask buyers to bid ${b}_{i}\in [0,1]$ on the item and the winner of the item is designating via an “allocation rule” that maps bids $b\in {[0,1]}^{n}$ to a winner in $\{0,...,n\}$ (0 refers to the no winner case). Then the payment rule $p:{[0,1]}^{n}\to {[0,1]}^{n}$ indicates the amount of money that each bidder must pay to the seller. Auctions are specific cases of a broader family of “mechanisms”. Knowing the allocation and payment rules, bidders have incentives to bid strategically. Different auctions (or rules) end up with different revenue to the seller, who can choose the optimal rules. This is rather standard in economics, but these interactions become way more intricate when repeated over time (as in the online ad market 106), when several items are sold at the same time (for instance in bundles), when the buyers have partial information about the actual value of the item 121 and/or reciprocally when the seller does not know the value distributions of the buyer. In that case, she might be tempted to try to learn them from the previous bids in order to design the optimal mechanism. Knowing this, the bidders have incentives to long term strategic behaviors, ending up in a quite complicated game between learning algorithms 107. This setting of interacting algorithms is actually of interest by itself, irrespectively of ad auctions. It is noteworthy also that traditional auction mechanisms do not guarantee any fairness notion and that the literature on fixing that (for applications where it matters) is only nascent 54, 105, 58, 80.
Matching 89, 104
A matching is nothing more than a bipartite graph between some agents (patients, doctors, students) and some resources (respectively, organs, hospital, schools). The objective is to figure out what is the “optimal” matching for a given criterion. Interestingly, there are two different—and mostly unrelated yet—concepts of “good matching”. The first one is called “stable” in the sense that each agent expresses preferences over resources (and viceversa) and be such that no couple (agentresource) that are unpaired would prefer to be paired together than with their current paired resource/agent. In the other concept of matching, agents and resources are described by some features (say vectors in ${R}^{d}$, denoted by ${a}_{n}$ for agents and ${r}_{m}$ for resources) and pairing ${a}_{n}$ to ${r}_{m}$ incurs a cost of $c({a}_{n},{r}_{m})$, for some a given function $c:{\left({R}^{d}\right)}^{2}\to [0,1]$. The objective is then to minimize the total cost of the matching ${\sum}_{n}c({a}_{n},{r}_{\sigma \left(n\right)})$, where $\sigma \left(n\right)$ is the resource allocated to agent $n$.
Matching is used is many different applications such as university admission (e.g., in Parcoursup). Notice that strategic interactions arise in matching if agents or resources can disclose their preferences/features to each other. Learning is also present as soon as not everything is known, e.g., the preferences or costs. Many applications of matching (again, such as college admission) are typical examples where fairness and privacy are of utmost importance. Finally, matching is also at the basis of several Internet applications and Criteo products, for instance to solve the problem of matching a given ad budget to fixed ad slots.
Ethical notions in those usecases
In both problems, individual users are involved and there is a clear need to consider fairness and privacy. However, the precise motivation and instantiation of these notions depends on the specific usecase. In fact, it is often part of the research question to decide which are the most relevant fairness and privacy notions, as mentioned in Section 2.1. We will throughout the team's life put an important focus on this question, as well as on the question of the impact of the chosen notion on performance.
4.2 Application areas
In FAIRPLAY, we consider both applications to Criteo usecases (online advertisement) and other applications (with other appropriate partners).
4.2.1 Online advertisement
Online advertising offers an application area for all of the research themes of FAIRPLAY; which we investigate primarily with Criteo.
First, online advertising is a typical application of online auctions and we consider applications of the work on auctions to Criteo usecases, in particular the work on advertisercentric fairness where the advertiser is Criteo. From a practical point of view, privacy will have to be enforced in such applications. For instance, when information is provided to advertisers to define audiences or to visualize the performance of their campaigns (insights) there is a possibility of leaking sensitive information on users. In particular, excellent proxies on protected attributes should probably not be leaked to advertisers, or transformed before (e.g., with the differential privacy techniques). This is therefore also an application of the fairnessvsprivacy research thread.
Note that, even before considering those questions, the first very important theoretical question is to determine what is the more appropriate definition of fairness (as there are, as mentioned above, many different variations) in those applications. We recall that it is wellknown that usual fairness metrics are not compatible 88. Moreover, in online advertising, fairness can be measured in term of bidding and recommendation or in term of what ads are actually displayed. Being fair on bidding does not lead to fairness in ads displaying 105, mainly because of the other advertising actors. While fairness in bidding and/or recommendation seem the most important because they only rely on our models, external auditors can easily get data on which ads we display.
We will also investigate applications of fair matching techniques to online advertsing and to Criteo matching products—namely retargeting (personalized ads displayed on a website) and retail media (sponsored products on a merchant website). Indeed, one of Criteo major products, retail media can be cast as an online matching problem. On a given ecommerce website (say, target), several advertisers—currently brands—are running campaigns so that their products are “sponsored” or “boosted”, i.e., they appear higher on the list of results of a given query. The budgets (from a Criteo perspective) must be cleared (daily, monthly or annually). This constraint is easy thanks to the high traffic, but the main issue is that, without control/pacing/matching in times, the budget is depleted after only a few hours on a relatively low quality traffic (i.e., users that generate few conversions hence a small ROI for the advertisers). The question is therefore whether an individual user should be matched or not to boosted/sponsored products at a given time so that the ROI of the advertisers is maximized, the campaign budget is depleted and the retailer does not suffer too much from this active corruption of its organic results. Those are three different and concurrent objectives (for respectively the advertisers, Criteo and the retailers) that must be somehow conciliated. This problem (and more generally this type of problems) offers a rich application area to the FAIRPLAY research program. Indeed, it is crucial to ensure that fairness and privacy are respected. On the other hand, users, clicks, conversion arrival are not “worst case”. They rather follow some complicated—but certainly learnable—process; which allows applying our results on exploiting structure.
4.2.2 “Physical matching”
We investigate a number of other applications of matching: assignment of daycare slots to kids, mutation of professors to different academies, assignment of kidneys to patients, assignment of job applicants to jobs. In all these applications, there are crucial constraints of fairness that complicate the matching. We leverage existing partnership with the CNAF, the French Ministry of Education and the Agence de la biomédecine in Paris for the first three applications; for the last we will consolidate a nascent partnership with Pole Emploi and LinkedIn.
5 Highlights of the year
The team's convention was officially signed by Criteo, ENSAE and Inria in December 2023. Simon Mauras was hired as a CRCN in 2023 and will join the team in early 2024. The team (Vianney Perchet and Julien Combe) organized a very successful weeklong conference at CIRM in December on “From matchings to markets. A tale of Mathematics, Economics and Computer Science.”
5.1 Awards
The team received a best paper award at ICML 2023 for the paper “Local and adaptive mirror descents in extensiveform games” by Côme Fiegel, Pierre Ménard, Tadashi Kozuno, Rémi Munos, Vianney Perchet, Michal Valko 19.
6 New results
6.1 Auctions and mechanism design
Participants: Benjamin Heymann, Simon Finster.
In 18, we consider the problem of maximizing the success probability of policy allocations in online bidding systems. The effectiveness of advertising in ecommerce largely depends on the ability of merchants to bid on and win impressions for their targeted users. The bidding procedure is highly complex due to various factors such as market competition, user behavior, and the diverse objectives of advertisers. In this paper we consider the problem at the level of user timelines instead of individual bid requests, manipulating full policies (i.e. predefined bidding strategies) and not bid values. In order to optimally allocate policies to users, typical multiple treatments allocation methods solve knapsacklike problems which aim at maximizing an expected value under constraints. In the industrial contexts such as online advertising, we argue that optimizing for the probability of success is a more suited objective than expected value maximization, and we introduce the SuccessProbaMax algorithm that aims at finding the policy allocation which is the most likely to outperform a fixed reference policy. Finally, we conduct comprehensive experiments both on synthetic and realworld data to evaluate its performance. The results demonstrate that our proposed algorithm outperforms conventional expectedvalue maximization algorithms in terms of success rate.
In 20, we study a mechanism design problem for pool testing. Largescale testing is crucial in pandemic containment, but resources are often prohibitively constrained. We study the optimal application of pooled testing for populations that are heterogeneous with respect to an individual's infection probability and utility that materializes if included in a negative test. We show that the welfare gain from overlapping testing over nonoverlapping testing is bounded. Moreover, nonoverlapping allocations, which are both conceptually and logistically simpler to implement, are empirically nearoptimal, and we design a heuristic mechanism for finding these nearoptimal test allocations. In numerical experiments, we highlight the efficacy and viability of our heuristic in practice. We also implement and provide experimental evidence on the benefits of utilityweighted pooled testing in a realworld setting. Our pilot study at a higher education research institute in Mexico finds no evidence that performance and mental health outcomes of participants in our testing regime are worse than under the firstbest counterfactual of full access for individuals without testing.
In 21, we study substitutes markets with budget constraints. Markets with multiple divisible goods have been studied widely from the perspective of revenue and welfare. In general, it is well known that envyfree revenuemaximal outcomes can result in lower welfare than competitive equilibrium outcomes. We study a market in which buyers have quasilinear utilities with linear substitutes valuations and budget constraints, and the seller must find prices and an envyfree allocation that maximise revenue or welfare. Our setup mirrors markets such as ad auctions and auctions for the exchange of financial assets. We prove that the unique competitive equilibrium prices are also envyfree revenuemaximal. This coincidence of maximal revenue and welfare is surprising and breaks down even when buyers have piecewiselinear valuations. We present a novel characterisation of the set of 'feasible' prices at which demand does not exceed supply, show that this set has an elementwise minimal price vector, and demonstrate that these prices maximise revenue and welfare. The proof also implies an algorithm for finding this unique price vector.
6.2 Online algorithms and learning
Participants: Matthieu Lerasle, Patrick Loiseau, Vianney Perchet, Dorian Baudry, Nadav Merlis.
6.2.1 Learning in games
In citefiegel:hal04416177, we study how to learn $\u03f5$optimal strategies in zerosum imperfect information games (IIG) with trajectory feedback. In this setting, players update their policies sequentially based on their observations over a fixed number of episodes, denoted by $T$. Existing procedures suffer from high variance due to the use of importance sampling over sequences of actions (Steinberger et al., 2020; McAleer et al., 2022). To reduce this variance, we consider a fixed sampling approach, where players still update their policies over time, but with observations obtained through a given fixed sampling policy. Our approach is based on an adaptive Online Mirror Descent (OMD) algorithm that applies OMD locally to each information set, using individually decreasing learning rates and a regularized loss. We show that this approach guarantees a convergence rate of $\tilde{O}\left({T}^{1/2}\right)$ with high probability and has a nearoptimal dependence on the game parameters when applied with the best theoretical choices of learning rates and sampling policies. To achieve these results, we generalize the notion of OMD stabilization, allowing for timevarying regularization with convex increments.
6.2.2 Bandits, control and reinforcement learning
In 10, we consider the problem of regret minimization in nonparametric stochastic bandits. When the rewards are known to be bounded from above, there exists asymptotically optimal algorithms, with asymptotic regret depending on an infimum of KullbackLeibler divergences (KL). These algorithms are computationally expensive and require storing all past rewards, thus simpler but nonoptimal algorithms are often used instead. We introduce several methods to approximate the infimum KL which reduce drastically the computational and memory costs of existing optimal algorithms, while keeping their regret guaranties. We apply our findings to design new variants of the MED and IMED algorithms, and demonstrate their interest with extensive numerical simulations.
In 17, we introduce Dynamic Contextual Markov Decision Processes (DCMDPs), a novel reinforcement learning framework for historydependent environments that generalizes the contextual MDP framework to handle nonMarkov environments, where contexts change over time. We consider special cases of the model, with a focus on logistic DCMDPs, which break the exponential dependence on history length by leveraging aggregation functions to determine context transitions. This special structure allows us to derive an upperconfidencebound style algorithm for which we establish regret bounds. Motivated by our theoretical results, we introduce a practical modelbased algorithm for logistic DCMDPs that plans in a latent space and uses optimism over historydependent features. We demonstrate the efficacy of our approach on a recommendation task (using MovieLens data) where user behavior dynamics evolve in response to recommendations.
In 4, we consider the diffusive limit of a typical purejump Markovian control problem as the intensity of the driving Poisson process tends to infinity. We show that the convergence speed is provided by the Hölder constant of the Hessian of the limit problem, and explain how correction terms can be constructed. This provides an alternative efficient method for the numerical approximation of the optimal control of a purejump problem in situations with very high intensity of jump. We illustrate this approach in the context of a display advertising auction problem.
6.2.3 Online (and offline) algorithms
In 15, we study online algorithms with advice querying under a budget constraint. Several problems have been extensively studied in the learningaugmented setting, where the algorithm has access to some, possibly incorrect, predictions. However, it is assumed in most works that the predictions are provided to the algorithm as input, with no constraint on their size. In this paper, we consider algorithms with access to a limited number of predictions, that they can request at any time during their execution. We study three classical problems in competitive analysis, the ski rental problem, the secretary problem, and the nonclairvoyant job scheduling. We address the question of when to query predictions and how to use them.
In 16, we also study online algorithms with predictions. A popular approach to go beyond the worstcase analysis of online algorithms is to assume the existence of predictions that can be leveraged to improve performances. Those predictions are usually given by some external sources that cannot be fully trusted. Instead, we argue that trustful predictions can be built by algorithms, while they run. We investigate this idea in the illustrative context of static scheduling with exponential job sizes. Indeed, we prove that algorithms agnostic to this structure do not perform better than in the worst case. In contrast, when the expected job sizes are known, we show that the best algorithm using this information, called FollowThePerfectPrediction (FTPP), exhibits much better performances. Then, we introduce two adaptive explorethencommit types of algorithms: they both first (partially) learn expected job sizes and then follow FTPP once their selfpredictions are confident enough. On the one hand, ETCU explores in "series", by completing jobs sequentially to acquire information. On the other hand, ETCRR, inspired by the optimal worstcase algorithm RoundRobin (RR), explores efficiently in "parallel". We prove that both of them asymptotically reach the performances of FTPP, with a faster rate for ETCRR. Those findings are empirically evaluated on synthetic data.
In 9, we study scheduling of moldable tasks, in the offline setting. Moldable tasks allow schedulers to determine the number of processors assigned to each task, thus enabling efficient use of largescale parallel processing systems. We consider the problem of scheduling independent moldable tasks on processors and propose a new perspective of the existing speedup models: as the number p of processors assigned to a task increases, the speedup is linear if $p$ is small and becomes sublinear after $p$ exceeds a threshold. Based on this, we propose an efficient approximation algorithm to minimize the makespan. As a byproduct, we also propose an approximation algorithm to maximize the sum of values of tasks completed by a deadline; this scheduling objective is considered for moldable tasks for the first time while similar works have been done for other types of parallel tasks.
6.2.4 Statistical estimation
In 11, we discuss an application of Stochastic Approximation to statistical estimation of highdimensional sparse parameters. The proposed solution reduces to resolving a penalized stochastic optimization problem on each stage of a multistage algorithm; each problem being solved to a prescribed accuracy by the nonEuclidean Composite Stochastic Mirror Descent (CSMD) algorithm. Assuming that the problem objective is smooth and quadratically minorated and stochastic perturbations are subGaussian, our analysis prescribes the method parameters which ensure fast convergence of the estimation error (the radius of a confidence ball of a given norm around the approximate solution). This convergence is linear during the first "preliminary" phase of the routine and is sublinear during the second "asymptotic" phase. We consider an application of the proposed approach to sparse Generalized Linear Regression problem. In this setting, we show that the proposed algorithm attains the optimal convergence of the estimation error under weak assumptions on the regressor distribution. We also present a numerical study illustrating the performance of the algorithm on highdimensional simulation data.
In 8, we study a changepoint detection problem. Given a times series $\mathbf{Y}$ in ${\mathbb{R}}^{n}$, with a piecewise contant mean and independent components, the twin problems of changepoint detection and changepoint localization respectively amount to detecting the existence of times where the mean varies and estimating the positions of those changepoints. In this work, we tightly characterize optimal rates for both problems and uncover the phase transition phenomenon from a global testing problem to a local estimation problem. Introducing a suitable definition of the energy of a changepoint, we first establish in the single changepoint setting that the optimal detection threshold is $\sqrt{2loglog\left(n\right)}$. When the energy is just above the detection threshold, then the problem of localizing the changepoint becomes purely parametric: it only depends on the difference in means and not on the position of the changepoint anymore. Interestingly, for most changepoint positions, including all those away from the endpoints of the time series, it is possible to detect and localize them at a much smaller energy level. In the multiple changepoint setting, we establish the energy detection threshold and show similarly that the optimal localization error of a specific changepoint becomes purely parametric. Along the way, tight optimal rates for Hausdorff and l 1 estimation losses of the vector of all changepoints positions are also established. Two procedures achieving these optimal rates are introduced. The first one is a leastsquares estimator with a new multiscale penalty that favours well spread changepoints. The second one is a twostep multiscale postprocessing procedure whose computational complexity can be as low as $O\left(nlog\right(n\left)\right)$. Notably, these two procedures accommodate with the presence of possibly many lowenergy and therefore undetectable changepoints and are still able to detect and localize highenergy changepoints even with the presence of those nuisance parameters.
In 6, we study variable selection, monotone likelihood ratio and group sparsity. In the pivotal variable selection problem, we derive the exact nonasymptotic minimax selector over the class of all ssparse vectors, which is also the Bayes selector with respect to the uniform prior. While this optimal selector is, in general, not realizable in polynomial time, we show that its tractable counterpart (the scan selector) attains the minimax expected Hamming risk to within factor 2, and is also exact minimax with respect to the probability of wrong recovery. As a consequence, we establish explicit lower bounds under the monotone likelihood ratio property and we obtain a tight characterization of the minimax risk in terms of the best separable selector risk. We apply these general results to derive necessary and sufficient conditions of exact and almost full recovery in the location model with light tail distributions and in the problem of group variable selection under Gaussian noise.
6.3 Privacy, Fairness, and Transparency
Participants: Cristina Butucea, Patrick Loiseau, Vianney Perchet.
In 14, we consider the problem of online allocation subject to a longterm fairness penalty. Contrary to existing works, however, we do not assume that the decisionmaker observes the protected attributes – which is often unrealistic in practice. Instead they can purchase data that help estimate them from sources of different quality; and hence reduce the fairness penalty at some cost. We model this problem as a multiarmed bandit problem where each arm corresponds to the choice of a data source, coupled with the online allocation problem. We propose an algorithm that jointly solves both problems and show that it has a regret bounded by $O\left(\sqrt{T}\right)$. A key difficulty is that the rewards received by selecting a source are correlated by the fairness penalty, which leads to a need for randomization (despite a stochastic setting). Our algorithm takes into account contextual information available before the source selection, and can adapt to many different fairness notions. We also show that in some instances, the estimates used can be learned on the fly.
In 12, we study the related problem of transparency, in the particular case of targeted advertising. Several targeted advertising platforms offer transparency mechanisms, but researchers and civil societies repeatedly showed that those have major limitations. In this paper, we propose a collaborative ad transparency method to infer, without the cooperation of ad platforms, the targeting parameters used by advertisers to target their ads. Our idea is to ask users to donate data about their attributes and the ads they receive and to use this data to infer the targeting attributes of an ad campaign. We propose a Maximum Likelihood Estimator based on a simplified Bernoulli ad delivery model. We first test our inference method through controlled ad experiments on Facebook. Then, to further investigate the potential and limitations of collaborative ad transparency, we propose a simulation framework that allows varying key parameters. We validate that our framework gives accuracies consistent with realworld observations such that the insights from our simulations are transferable to the real world. We then perform an extensive simulation study for ad campaigns that target a combination of two attributes. Our results show that we can obtain good accuracy whenever at least ten monitored users receive an ad. This usually requires a few thousand monitored users, regardless of population size. Our simulation framework is based on a new method to generate a synthetic population with statistical properties resembling the actual population, which may be of independent interest.
In 13, we also study a transparency problem related to fairness, but in the context of decentralized systems. In permissionless blockchains, transaction issuers include a fee to incentivize miners to include their transactions. To accurately estimate this prioritization fee for a transaction, transaction issuers (or blockchain participants, more generally) rely on two fundamental notions of transparency, namely contention and prioritization transparency. Contention transparency implies that participants are aware of every pending transaction that will contend with a given transaction for inclusion. Prioritization transparency states that the participants are aware of the transaction or prioritization fees paid by every such contending transaction. Neither of these notions of transparency holds well today. Private relay networks, for instance, allow users to send transactions privately to miners. Besides, users can offer fees to miners via either direct transfers to miners' wallets or offchain paymentsneither of which are public. In this work, we characterize the lack of contention and prioritization transparency in Bitcoin and Ethereum resulting from such practices. We show that private relay networks are widely used and private transactions are quite prevalent. We show that the lack of transparency facilitates miners to collude and overcharge users who may use these private relay networks despite them offering little to no guarantees on transaction prioritization. The lack of these transparencies in blockchains has crucial implications for transaction issuers as well as the stability of blockchains. Finally, we make our data sets and scripts publicly available.
In 5, we address the problem of variable selection in a highdimensional but sparse mean model, under the additional constraint that only privatised data are available for inference. The original data are vectors with independent entries having a symmetric, strongly logconcave distribution on $\mathbb{R}$. For this purpose, we adopt a recent generalisation of classical minimax theory to the framework of local $\alpha $differential privacy. We provide lower and upper bounds on the rate of convergence for the expected Hamming loss over classes of at most $s$sparse vectors whose nonzero coordinates are separated from 0 by a constant $a>0$. As corollaries, we derive necessary and sufficient conditions (up to log factors) for exact recovery and for almost full recovery. When we restrict our attention to noninteractive mechanisms that act independently on each coordinate our lower bound shows that, contrary to the nonprivate setting, both exact and almost full recovery are impossible whatever the value of $a$ in the highdimensional regime such that $n{\alpha}^{2}/{d}^{2}\le 1$. However, in the regime $n{\alpha}^{2}/{d}^{2}\gg log\left(d\right)$ we can exhibit a critical value ${a}^{*}$ (up to a logarithmic factor) such that exact and almost full recovery are possible for all $a\gg {a}^{*}$ and impossible for $a\le {a}^{*}$. We show that these results can be improved when allowing for all noninteractive (that act globally on all coordinates) locally $\alpha $differentially private mechanisms in the sense that phase transitions occur at lower levels.
In 7, we study interactive versus noninteractive locally differentially private estimation for the quadratic functional. Local differential privacy has recently received increasing attention from the statistics community as a valuable tool to protect the privacy of individual data owners without the need of a trusted third party. Similar to the classical notion of randomized response, the idea is that data owners randomize their true information locally and only release the perturbed data. Many different protocols for such local perturbation procedures can be designed. In most estimation problems studied in the literature so far, however, no significant difference in terms of minimax risk between purely noninteractive protocols and protocols that allow for some amount of interaction between individual data providers could be observed. In this paper we show that for estimating the integrated square of a density, sequentially interactive procedures improve substantially over the best possible noninteractive procedure in terms of minimax rate of estimation. In particular, in the noninteractive scenario we identify an elbow in the minimax rate at $s=3/4$, whereas in the sequentially interactive scenario the elbow is at $s=1/2$. This is markedly different from both, the case of direct observations, where the elbow is well known to be at $s=1/4$, as well as from the case where Laplace noise is added to the original data, where an elbow at $s=9/4$ is obtained. We also provide adaptive estimators that achieve the optimal rate up to logfactors, we draw connections to nonparametric goodnessoffit testing and estimation of more general integral functionals and conduct a series of numerical experiments. The fact that a particular locally differentially private, but interactive, mechanism improves over the simple noninteractive one is also of great importance for practical implementations of local differential privacy.
7 Partnerships and cooperations
7.1 International research visitors
7.1.1 Visits of international scientists
Other international visits to the team
A. Rohde
 Professor
 Freiburg University
 Germany

Dates:
May 31  June 1, 2023

Context of the visit:
Research stay
A. Meister
 Professor
 Rostock University
 Germany

Dates:
May 31  June 1, 2023

Context of the visit:
Research stay
A. Celli
 Professor
 Bocconi University
 Italy

Dates:
Sept 11 Sept 22, 2023

Context of the visit:
Research stay
7.1.2 Visits to international teams
Research stays abroad
Cristina Butucea

Visited institution:
Heidelberg University

Country:
UK

Dates:
February 2024, 2023

Context of the visit:
Research, discussions

Visited institution:
Nottingham University

Country:
UK

Dates:
July 27, 2023

Context of the visit:
Research, discussions
Vianney Perchet

Visited institution:
MIT

Country:
US

Dates:
May 1  May 7, 2023

Context of the visit:
Research, discussions & seminar
7.2 National initiatives
Foundry (PEPR IA)
Participants: Patrick Loiseau.

Title:
Foundry: Foundation of robustness and reliability in AI

Partner Institution(s):
 Inria
 CNRS
 Université Paris Dauphine
 Institut Mines Telecom
 ENS de Lyon

Date/Duration:
20232027 (4 years)

Additionnal info/keywords:
PEPR IA projet cible, 245k euros. Fairness, matching, auctions.
FairPlay (ANR JCJC)
Participants: Patrick Loiseau.

Title:
FairPlay: Fair algorithms via game theory and sequential learning

Partner Institution(s):
 Inria

Date/Duration:
20212025 (4 years)

Additionnal info/keywords:
ANR JCJC project, 245k euros. Fairness, matching, auctions.
Explainable and Responsible AI (MIAI chair)
Participants: Patrick Loiseau.

Title:
Explainable and Responsible AI chair of the MIAI @ Grenoble Alpes institute

Partner Institution(s):
 Univ. Grenoble Alpes

Date/Duration:
20192023 (4 years)

Additionnal info/keywords:
Chair of the MIAI @ Grenoble Alpes institute coheld by Patrick Loiseau. Fairness, privacy.
BOLD (ANR)
Participants: Vianney Perchet.

Title:
BOLD: Beyond Online Learning for Better Decisions

Partner Institution(s):
 Crest, Genes

Date/Duration:
20192024 (4.5 years)

Additionnal info/keywords:
ANR project, 270k euros. online learning, optimization, bandits.
8 Dissemination
8.1 Promoting scientific activities
8.1.1 Scientific events: organisation
General chair, scientific chair
Participants: Vianney Perchet.

Title:
From matchings to markets. A tale of Mathematics, Economics and Computer Science.

Partner Institution(s):
 Crest, Genes

Date/Duration:
December 2023

Location:
CIRM, Marseille
8.1.2 Scientific events: selection
Member of the conference program committees

Patrick Loiseau:
NeurIPS, ECMLPKDD, EWAF

Vianney Perchet:
NeurIPS, ICLR, ICML, COLT, ALT

Hugo Richard:
NeurIPS, ICML, AISTATS

Marc Abeille:
ALT

Clément Calauzènes:
ICML, NeurIPS

Benjamin Heymann
NeurIPS, AISTATS

Maxime Vono
NeurIPS
8.1.3 Journal
Member of the editorial boards

Vianney Perchet:
Foundations and Trends in Machine Learning, Operation Research, Operation Research Letters, Journal of Machine Learning Research, Journal of Dynamics and Games,

Cristina Butucea:
Annals of Statistics, Bernoulli
Reviewer  reviewing activities

Patrick Loiseau:
Journal of Machine Learning Research, Mathematics of Operation Research, Games and Economic Behavior

Vianney Perchet:
Annals of Statistics, Mathematics of Operation Research, Journal of the ACM

Matthieu Lerasle
Annals of statistics, Journal of the European Mathematical Society, Probability and Related Fields, Journal of Machine Learning Research, Journal of the American Statistical Association.

Marc Abeille:
Journal of Machine Learning Research

Maxime Vono
Journal of Computational and Graphical Statistics

Benjamin Heymann:
SIAM Control and Optimization, Internation Journal of Game Theory, Mathematical Reviews
8.1.4 Invited talks

Patrick Loiseau:
Alpine Game Theory Symposium (Grenoble), Columbia University, Ethics of Public Robots and AI (Skema)

Vianney Perchet:
Alpine Game Theory Symposium (Grenoble), Algorithms, Learning, and Games (Sicily), Optimization and Statistical Learning (les Houches), Mathematics of Data Science (Singapore), Seminar of the Maths Department of MIT (Boston), Games Learning and Networks (Singapore), Optimization 2023 (Seattle), Workshop on Information & Learning in Decisions & Operations (INSEAD), FILOFOCS 2023 (Paris), R. Cominetti's Feist (Chile), NeurIPS Workshop I cannot Believe It's Not Better (New Orleans), New Methods in Statistics (Marseille)

Matthieu Lerasle
Journées Statistiques du Sud, Journées ALEA, Colloquium Université Rouen, Séminaire Université du Luxembourg, Séminaire Geneva School of Economics and Management.

Cristina Butucea
University of Vienna, Séminaire de statistique IHP, Anniversary Conference M. Neumann University of Bamberg, Workshop on Tests and Bandits University of Potsdam

Marc Abeille:
UpperBound conference (2023), Reinforcement Learning Theory seminars

Benjamin Heymann:
Fundamental Challenges in Causality seminar workshop, Causality in Practice workshop, Causality Discussion Group

Maxime Vono:
Federated Learning One World Seminar
8.1.5 Scientific expertise

Vianney Perchet:
Expert for the evaluation of the LABEX MME:DII

Cristina Butucea:
Reviewer for tenure committee Harvard University, hiring committees France
8.2 Teaching  Supervision  Juries
8.2.1 Supervision

Patrick Loiseau:
PhD students: Rémi Castera, Mathieu Molina, Mélissa Tamine; postdocs: Felipe Garrido Lucero, Simon Finster, Denis Sokolov

Vianney Perchet:
PhD students: Sasila Ilandarideva, Flore Sentenac, Come Fiegel, Maria Cherifa, Mathieu Molina, Ziyad Benomar, Mike Liu, Hafedh El Ferchichi. postdocs: Felipe Garrido Lucero, Nadav Merlis, Dorian Baudry

Matthieu Lerasle
PhD Students: Clara Carlier, Hugo Chardon, Hafedh El Ferchichi.

Cristina Butucea
PhD students: Nayel Bettache, Henning Stein

Marc Abeille:
Lorenzo Croissant

Clément Calauzènes:
Morgane Goibert, Maria Cherifa

Benjamin Heymann:
Mélissa Tamine

Maxime Vono:
Mélissa Tamine
8.2.2 Juries

Patrick Loiseau:
PhD jury A. Bardou (reviewer), V. Do (reviewer)

Vianney Perchet:
PhD jury: J. Achddou (reviewer), D. Beaudry, A. Bismuth, C.S. Gauthier (reviewer), H. Dakdouk (reviewer), S. Gaucher, F. Hu, G. Rizk, P. Muller, HDR jury: A. Simonetto (reviewer)

Matthieu Lerasle
PhD Jury: S. ArradiAlaoui (reviewer), J. Cheng. HDR Jury: P. Mozharovskyi, A, Sabourin (internal reviewer).

Cristina Butucea
PhD Jury: E. Pilliat (Montpellier), C Deslandes (CMAP)
8.3 Teaching

ENSAE:

Advanced Optimization
Third year, lectures

Algorithm Design and Analysis
Third year, lectures

Theoretical Foundations of Machine Learning
Second year, lectures

Stopping time and online algorithms
Third year, lectures

Statistics
(ML) 1st and second year

Nonparametric Statistics
3rd year, M2

Mathematical Foundations of Probabilities
1st year

Programming project
1st year

Advanced Optimization

Ecole Polytechnique:

INF421: design and analysis of algorithms
(Patrick Loiseau). Secondyear level, PCs.

INF581: Advanced Machine Learning and Autonomous Agents
(Patrick Loiseau). Thirdyear/M1 level, lectures and labs.

MAP433: Statistics
(ML). Firstyear cycle polytechnicien, PCs.

MAP576: Learning Theory
(ML). Secondyear cycle polytechnicien, Lecture.

INF421: design and analysis of algorithms

Université ParisSaclay:

High Dimensional Probability
(ML). Master 2

Stopping Time and Random Algorithm
(ML). Master 2

High Dimensional Probability

PSL:

Introduction to machine learning
(Hugo Richard). L3 level, Lectures and labs.

Introduction to machine learning

Master IASD:

Recommender Systems
(Clément Calauzènes). Master 2, Lectures.

Recommender Systems
9 Scientific production
9.1 Major publications
 1 inproceedingsLocal and adaptive mirror descents in extensiveform games.2023 International Conference on Machine LearningNew Orleans, United StatesSeptember 2023HAL
 2 inproceedingsTradingoff price for data quality to achieve fair online allocation.NeurIPS 2023  37th Conference on Neural Information Processing SystemsNew orleans, USA, United States2023, 143HAL
 3 articleOptimal ChangePoint Detection and Localization.Annals of Statistics5142023, 15861610HALDOI
9.2 Publications of the year
International journals
 4 articleDiffusive limit approximation of purejump optimal stochastic control problems.Journal of Optimization Theory and ApplicationsJanuary 2023HALDOIback to text
 5 articlePhase transitions for support recovery under local differential privacy.Mathematical Statistics and Learning61June 2023, 150HALDOIback to text
 6 articleVariable selection, monotone likelihood ratio and group sparsity.The Annals of Statistics511February 2023HALDOIback to text
 7 articleInteractive versus noninteractive locally differentially private estimation: Two elbows for the quadratic functional.The Annals of Statistics512April 2023HALDOIback to text
 8 articleOptimal ChangePoint Detection and Localization.Annals of Statistics5142023, 15861610HALDOIback to text
 9 articleEfficient approximation algorithms for scheduling moldable tasks.European Journal of Operational Research3101October 2023, 7183HALDOIback to text
International peerreviewed conferences
 10 inproceedingsFast Asymptotically Optimal Algorithms for NonParametric Stochastic Bandits.Advances in Neural Information Processing Systems 36 (NeurIPS 2023)Thirtyseventh Conference on Neural Information Processing SystemsNew Orleans (Louisiana), United StatesDecember 2023HALback to text
 11 inproceedingsStochastic Mirror Descent for LargeScale Sparse Recovery.26th International Conference on Artificial Intelligence and Statistics (AISTATS)Valencia, SpainApril 2023HALback to text
 12 inproceedingsCollaborative Ad Transparency: Promises and Limitations.SP 2023  44th IEEE Symposium on Security and PrivacySan Francisco, United StatesMay 2023HALback to text
 13 inproceedingsDissecting Bitcoin and Ethereum Transactions: On the Lack of Transaction Contention and Prioritization Transparency in Blockchains.Proceedings of Financial Cryptography and Data Security 2023FC 2023  Financial Cryptography and Data Security 2023Bol, Brač, CroatiaMay 2023HALback to text
 14 inproceedingsTradingoff price for data quality to achieve fair online allocation.NeurIPS 2023  37th Conference on Neural Information Processing SystemsNew orleans, USA, United States2023, 143HALback to text
 15 inproceedingsAdvice Querying under Budget Constraint for Online Algorithms.NeurIPS 2023  37th Conference on Neural Information Processing SystemsNew Orleans, United StatesDecember 2023HALback to text
 16 inproceedingsStatic Scheduling with Predictions Learned through Efficient Exploration.2023 International Conference on Machine LearningHonolulu (Hawai), United StatesMay 2022HALback to text
 17 inproceedingsReinforcement Learning with HistoryDependent Dynamic Contexts.ICMLHonolulu, United StatesMay 2023HALback to text
Conferences without proceedings
 18 inproceedingsMaximizing the Success Probability of Policy Allocations in Online Systems.AAAI2024Vancouver, CanadaarXiv2023HALDOIback to text
 19 inproceedingsBest paperLocal and adaptive mirror descents in extensiveform games.2023 International Conference on Machine LearningNew Orleans, United StatesSeptember 2023HALback to text
 20 inproceedingsWelfareMaximizing Pooled Testing.EC 2023  The 24th ACM Conference on Economics and ComputationLondon, United Kingdom2023HALback to text
 21 inproceedingsSubstitutes markets with budget constraints: solving for competitive and optimal prices.WINE 2023  The 19th Conference On Web And InterNet EconomicsShanghai, China2023HALback to text
Scientific book chapters
Edition (books, proceedings, special issue of a journal)
 23 proceedingsD.Denis BelomestnyC.Cristina ButuceaE.Enno MammenE.Eric MoulinesM.Markus ReißV.Vladimir UlyanovFoundations of Modern Statistics.Festschrift in Honor of Vladimir Spokoiny425Springer Proceedings in Mathematics & StatisticsSpringer International Publishing; Springer International Publishing2023HALDOI
Reports & preprints
 24 miscMultiArmed Bandits with Guaranteed Revenue per Arm.January 2024HAL
 25 miscAddressing bias in online selection with limited budget of comparisons.November 2023HAL
 26 miscTwosided Matrix Regression.March 2023HAL
 27 miscSimultaneous offthegrid learning of mixtures issued from a continuous dictionary.January 2024HAL
 28 miscStatistical Discrimination in Stable Matching.April 2023HAL
 29 miscNearcontinuous time Reinforcement Learning for continuous stateaction spaces.September 2023HAL
 30 miscAdapting to game trees in zerosum imperfect information games.December 2022HAL
 31 miscDUShapley: A Shapley Value Proxy for Efficient Dataset Valuation.June 2023HAL
 32 miscOnline Matching in Geometric Random Graphs.October 2023HAL
9.3 Cited publications
 33 unpublished24 CFR § 100.75  Discriminatory advertisements, statements and notices.., https://www.law.cornell.edu/cfr/text/24/100.75back to text
 34 articleCommunity detection and stochastic block models: recent developments.The Journal of Machine Learning Research1812017, 64466531back to text
 35 inproceedingsDiscrimination through optimization: How Facebook's ad delivery can lead to skewed outcomes.CSCW2019back to text
 36 inproceedingsOnline learning with feedback graphs: Beyond bandits.Annual Conference on Learning Theory40Microtome Publishing2015back to text
 37 articleOn matching and thickness in heterogeneous dynamic markets.Operations Research6742019, 927949back to text
 38 miscKidney Exchange in Dynamic Sparse Heterogenous Pools.2013back to text
 39 inproceedingsOnline stochastic optimization in the large: Application to kidney exchange.TwentyFirst International Joint Conference on Artificial Intelligence2009back to text
 40 inproceedingsImproved bounds for online stochastic matching.European Symposium on AlgorithmsSpringer2010, 170181back to text
 41 inproceedingsObtaining fairness using optimal transport theory. arXiv:1806.031952018, 125back to text
 42 articleOnline bipartite matching made simple.Acm Sigact News3912008, 8087back to text
 43 articleThe width of random graph orders.The Mathematical Scientist2001 1995back to text
 45 bookOnline Computation and Competitive Analysis.Cambridge University Presss1998back to text
 46 miscAn Experimental Study of Algorithms for Online Bipartite Matching.2018back to text
 47 inproceedingsSICMMAB: Synchronisation Involves Communication in Multiplayer MultiArmed Bandits.arXiv:1809.081512018, 131back to text
 48 inproceedingsUtility/Privacy Tradeoff through the lens of Optimal Transport.Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics108Proceedings of Machine Learning ResearchOnlinePMLRAugust 2020, 591601back to textback to text
 49 articleRegret Analysis of Stochastic and Nonstochastic Multiarmed Bandit Problems.Machine Learning512012, 1122back to text
 50 articleBounded regret in stochastic multiarmed bandits.Journal of Machine Learning Research: Workshop and Conference Proceedings (COLT)302013, 122134back to text
 51 inproceedingsOnline PrimalDual Algorithms for Maximizing AdAuctions Revenue.Berlin, Heidelberg2007, 253264back to textback to text
 52 articleLocal differential privacy: Elbow effect in optimal density estimation and adaptation over Besov ellipsoids.Bernoulli2632020, 17271764back to text
 53 articleInteractive versus noninteractive locally, differentially private estimation: Two elbows for the quadratic functional.Annals of Statsto appear2023back to text
 54 inproceedingsToward Controlling Discrimination in Online Ad Auctions.ICML2019back to textback to text
 55 inproceedings Regret minimization for reserve prices in secondprice auctions.Proceedings of SODA 20132013back to text
 56 bookPrediction, Learning, and Games.Cambridge University Press2006back to text
 57 inproceedingsCapacity bounded differential privacy.Advances in Neural Information Processing Systems2019, 34693478back to text
 58 miscFairness in ad auctions through inverse proportionality.arXiv:2003.139662020back to textback to text
 59 inproceedingsRandomized online matching in regular graphs.Proceedings of the TwentyNinth Annual ACMSIAM Symposium on Discrete AlgorithmsSIAM2018, 960979back to text
 60 inproceedingsPacing equilibrium in firstprice auction markets.EC2019back to text
 61 inproceedingsMultiplicative pacing equilibria in auction markets.WINE2018back to text
 62 inproceedingsProphet inequalities for iid random variables from an unknown distribution.Proceedings of the 2019 ACM Conference on Economics and Computation2019, 317back to text
 63 inproceedingsOn the Compatibility of Privacy and Fairness.FairUMAP2019back to text
 64 miscAmazon scraps secret AI recruiting tool that showed bias against women.Reuters, https://www.reuters.com/article/usamazoncomjobsautomationinsight/amazonscrapssecretairecruitingtoolthatshowedbiasagainstwomenidUSKCN1MK08G2018back to text
 65 articleDifferential privacy.Encyclopedia of Cryptography and Security2011, 338340back to text
 66 inproceedingsFairness through awareness.ITCS2012back to text
 67 inproceedingsCalibrating noise to sensitivity in private data analysis.Theory of cryptography conferenceSpringer2006, 265284back to text
 68 techreportOptimal PrivacyConstrained Mechanisms.C.E.P.R. Discussion Papers2019back to text
 69 inproceedingsOn the Effect of Positive Discrimination on Multistage Selection Problems in the Presence of Implicit Variance.EC2020back to text
 70 miscOnline Stochastic Matching: Beating 11/e.2009back to text
 71 techreportFairness in Precision Medicine.Data & Society2018back to text
 72 bookGame Theory.MIT press1991back to text
 73 articleLocal Differentially Private Regret Minimization in Reinforcement Learning.arXiv preprint arXiv:2010.077782020back to text
 74 articleLinear Regression from Strategic Data Sources.ACM Transactions on Economics and Computation82May 2020, 10:110:24back to text
 75 inproceedingsA Refined Mean Field Approximation.SIGMETRICS2017back to text
 76 articleProxy fairness.arXiv preprint arXiv:1806.112122018back to text
 77 inproceedingsBias in online freelance marketplaces: Evidence from taskrabbit and fiverr.CSCW2017back to text
 78 inproceedingsEquality of Opportunity in Supervised Learning.NIPS2016back to textback to text
 79 inproceedingsOnline task assignment in crowdsourcing markets.Twentysixth AAAI conference on artificial intelligence2012back to textback to text
 80 inproceedingsMultiCategory Fairness in Sponsored Search Auctions.FAT*2020back to textback to text
 81 articleMean field equilibria of dynamic auctions with learning.Management Science60122014, 29492970back to text
 82 articleOnline stochastic matching: New algorithms with better bounds.Mathematics of Operations Research3932014, 624646back to text
 83 articleOn the consistency of supervised learning with missing values.arXiv preprint arXiv:1902.069312019back to text
 84 articleAssessing algorithmic fairness with unobserved protected class using data combination.arXiv preprint arXiv:1906.002852019back to text
 85 inproceedingsOnline bipartite matching with unknown distributions.Proceedings of the fortythird annual ACM symposium on Theory of computing2011, 587596back to text
 86 inproceedingsAn optimal algorithm for online bipartite matching.Proceedings of the twentysecond annual ACM symposium on Theory of computing1990, 352358back to text
 87 inproceedingsAvoiding discrimination through causal reasoning.Advances in Neural Information Processing Systems2017, 656666back to text
 88 articleModelAgnostic Characterization of Fairness Tradeoffs.arXiv preprint arXiv:2004.034242020back to textback to text
 89 bookStable Marriage and Its Relation to Other Combinatorial Problems: An Introduction to the Mathematical Analysis of Algorithms.English translation, (CRM Proceedings and Lecture Notes), American Mathematical Society1996back to text
 90 bookAuction Theory.Elsevier2009back to text
 91 inproceedingsCounterfactual fairness.Advances in neural information processing systems2017, 40664076back to text
 92 miscThe long road to fairer algorithms.2020back to text
 93 articleAlgorithmic Bias? An Empirical Study of Apparent GenderBased Discrimination in the Display of STEM Career Ads.Management Science2019back to textback to text
 94 miscHow We Analyzed the COMPAS Recidivism Algorithm.ProPublica, https://www.propublica.org/article/howweanalyzedthecompasrecidivismalgorithm2016back to text
 95 inproceedingsCompeting Bandits in Matching Markets.arXiv:1906.053632019, 115back to text
 96 articleBusiness cycle dynamics under rational inattention.The Review of Economic Studies8242015, 15021532back to text
 97 inproceedingsOnline bipartite matching with random arrivals: an approach based on strongly factorrevealing lps.Proceedings of the fortythird annual ACM symposium on Theory of computing2011, 597606back to text
 98 articleOnline stochastic matching: Online actions based on offline statistics.Mathematics of Operations Research3742012, 559573back to textback to text
 99 bookMicroeconomic Theory.Oxford University Press1995back to text
 100 articleGreedy online bipartite matching on random graphs.arXiv preprint arXiv:1307.25362013back to text
 101 articleRational inattention to discrete choices: A new foundation for the multinomial logit model.American Economic Review10512015, 27298back to text
 102 articleOnline Matching and Ad Allocation.Found. Trends Theor. Comput. Sci.84October 2013, 265–368URL: https://doi.org/10.1561/0400000057DOIback to textback to textback to text
 103 bookRepeated Games.Econometric Society MonographsCambridge University Press2015back to text
 104 articleAlgorithms for the Assignment and Transportation Problems.Journal of the Society for Industrial and Applied Mathematics511957, 3238back to text
 105 inproceedingsBidding Strategies with Gender Nondiscrimination Constraints for Online Ad Auctions.FAT*2020back to textback to text
 106 inproceedingsThe bidder’s standpoint : a simple way to improve bidding strategies in revenuemaximizing auctions.Workshop on Learning in the Presence of Strategic Behavior (EC 2019)2019back to text
 107 inproceedingsLearning to bid in revenuemaximizing auctions.Proceedings of the 36th International Conference on Machine Learning (ICML 2019)2019, 47814789back to text
 108 bookAlgorithmic Game Theory.New York, NY, USACambridge University Press2007back to text
 109 articleA differential game on Wasserstein space. Application to weak approachability with partial monitoring.Journal of Dynamics and Games62019, 6585back to text
 110 articleThe multiarmed bandit problem with covariates.Annals of Statistics412013, 693721back to text
 111 bookElements of causal inference.The MIT Press2017back to text
 112 bookComputational Optimal Transport.ArXiv:1803.005672018back to text
 113 inproceedingsFair Inputs and Fair Outputs: The Incompatibility of Fairness in Privacy and Accuracy.FairUMAP2020back to text
 114 inproceedingsDistance makes the types grow stronger: a calculus for differential privacy.ACM Sigplan Notices452010, 157168back to text
 115 articleGeometrizing rates of convergence under local differential privacy constraints.Annals of Statistics4852020, 26462670back to text
 116 articleWhat's in a Name? Reducing Bias in Bios without Access to Protected Attributes.arXiv preprint arXiv:1904.052332019back to text
 117 articleAlgorithms that" Don't See Color": Comparing Biases in Lookalike and Special Ad Audiences.arXiv preprint arXiv:1912.075792019back to text
 118 articleImplications of rational inattention.Journal of monetary Economics5032003, 665690back to text
 119 inproceedingsOn the Potential for Discrimination in Online Targeted Advertising.ACM FAT*2018back to text
 120 bookTopics in optimal transportation.58Graduate studies in Mathematics, AMS2003back to text
 121 inproceedingsOnline learning in repeated auctions.Proceedings of COLT 20162016back to textback to text
 122 inproceedingsLearning NonDiscriminatory Predictors.COLT2017back to text
 123 inproceedingsCounterfactual Fairness: Unidentification, Bound and Algorithm..IJCAI2019, 14381444back to text
 124 inproceedingsPcfairness: A unified framework for measuring causalitybased fairness.Advances in Neural Information Processing Systems2019, 34043414back to text
 125 inproceedingsFairness Beyond Disparate Treatment & Disparate Impact: Learning Classification Without Disparate Mistreatment.WWW2017back to text
 126 inproceedingsFrom Parity to Preferencebased Notions of Fairness in Classification.NIPS2017back to text
 127 inproceedingsLearning Fair Representations.ICML2013back to text