2.1.1 Multi-agent 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 multi-agent model.

Each agent in a multi-agent 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 105. 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 multi-agent 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” 114: the ratio of the maximum social welfare to the (worst/best) social welfare when agents optimize individually.

The natural language to model and study multi-agent systems is game theory—see below for a list of tools and techniques on which FAIRPLAY relies, game theory being the first of them. Multi-agent 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 multi-agent systems involving users either as data providers or as individuals affected by the ML agent decision (or both).

Game theory and economics

Game theory 78 is the natural mathematical tool to model multiple interacting decision-makers (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 so-called 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” (one-shot) solution concept, but there also exist dynamic solution concepts for repeated games 62, 109.

Online and reinforcement learning 55

In online learning (a.k.a. multi-armed bandit 56, 116), 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 $ℱ$, i.e., such that the loss ${\sum }_{t=1}^{T}1\{Y_t\ne {\widehat{Y}}_t\}$ is $o\left(T\right)$-close to ${max}_{f\in ℱ}{\sum }_{t=1}^{T}1\{Y_t\ne f\left(X_t\right)\}$; the difference between those terms is called regret. The more general model with an underlying state of the world $S_t\in 𝒮$ 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 well-understood 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 53, 101.

Online algorithms 51 and theoretical computer science

Online algorithms are closely related to online learning with a major twist. In online learning, the agent has “0-look 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 “1-look 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 68,

Optimal transport 126

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 $𝒳$, the optimal transport problem consist in finding (if it exists, otherwise the problem can be relaxed) a transport map $T:𝒳\to 𝒳$ that minimizes ${\int }_{𝒳}c(x,T\left(x\right))d\mu \left(x\right)$ for some cost function $c:{𝒳}^2\to R$, under the constraint that $T♯\mu =\nu$, where $T♯\mu$ is the push-forward 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 115 thanks to its application to images and to new techniques to compute approximate matchings in a tractable way 118. Even more unexpected applications of optimal transport have been discovered: to protect privacy 54, fairness 47, 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.

3.1.1 Fairness in auction-based 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 self-interested 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 auction-based system (note that we focus on online ad auctions but the work has broader applicability).

• Advertiser-centric (or bidder-centric) fairness

  We first focus on advertiser-centric fairness, i.e., the advertiser of a third-party 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 mean-field games approach similar to 87 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 mean-field approximations 81 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.

• Platform-centric (or auction-centric) 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 67, 66, but using different multipliers for the different groups (i.e., different values of the sensitive attribute).

  Following recent theoretical work on auction fairness 60, 86, 64 (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 75, 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 decision-makers
  Next, we investigate cases with multiple decision-makers. We propose two cases in particular. The first one is the simple two-stage selection problem but where the decision-maker doing the first-stage selection is different from the decision-maker doing the second-stage 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 multiple-decision 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 regret-minimizing 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 two-sided 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 trade-off 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 long-term, 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 sub-groups defined by all intersections; this combinatorial is challenging and, for the moment, still exploratory.

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 108, 85, 45, 44. Recall that given a graph $G=(𝒱,ℰ)$—where $𝒱$ is a set of vertices and $ℰ$ is a set of edges—, a matching $ℳ\subset ℰ$ is a subset of edges such that each vertex belongs to at most one edge $e\in ℳ$. In that context, a perfect matching $ℳ$ is a matching where each vertex $v\in 𝒱$ is associated to an edge $e\in ℳ$, 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 108, 57. Consider a bipartite graph, where $𝒱=U\cup V$ 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 two-sided markets besides ad allocations such as assigning tasks to workers 85.

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 108. The seminal work 92 introduced an optimal algorithm for the adversarial case 48, that guarantees a CR of $1-\frac{1}{e}$; but focusing on a pessimistic worst-case. 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 $𝒟$88, 46, 76, 91, 103, 104. The classical approach consists in optimizing the CR over the distribution $𝒟$. Even in this seemingly optimistic framework, however, it is now known that there is no hope for a CR of more than 0.823 104. Moreover, this generally leads to very large linear programs (LP).

A more recent approach restricts the distribution $𝒟$ over which the problem is optimized to classes of graphs with an underlying stochastic structure. The benefit of this approach is two-fold: it gives hope for higher competitive ratios, and for simpler algorithms. Experiments also proved that complex algorithms optimized on $𝒟$ fared no better than simple greedy heuristics on “real-life” graphs 52. A few results along these lines show that is a promising path. For instance, 57 studied the problem on graphs where each vertex has degree at least $d$ and found a competitive ratio of $1-{(1-1/d)}^d$. On d-regular graphs, 65 designed a $1-O(\sqrt{logd}/\sqrt{d})$ competitive algorithm. 106 showed that greedy algorithms were highly efficient on Erdös-Renyi random graphs, with a competitive ratio of 0.837 in the worst case. 43 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 40 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 Erdös-Renyi model, where the inplace vertices $u\in 𝒰$ 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 Karp-Sipser algorithm on the graph 49. The second one involves the rich theory of graph weak local convergence 50. 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 bi-partite case). Each vertex $u\in 𝒰$ has a number of half-edges drawn for the same distribution ${\pi }_{𝒰}$ and each vertex $v\in 𝒱$ has a number of half-edges drawn from ${\pi }_V$ (with the assumption that the expected total numbers of half edges from $𝒰$ and $𝒱$ are the same). Then a vertex $v\in 𝒱$ that arrives in the sequential fashion has its half-edges “completed” by a (still free) half-edge of $𝒰$. This is a standard way of creating random graphs with (almost) fixed distribution of degrees. Here the question would simply be the competitive ratio of some greedy algorithm, whether the distributions ${\pi }_{𝒰}$ and ${\pi }_{𝒱}$ are known beforehand or learned on the fly. An interesting variant of this problem would be to assume the existence of a (hidden or not) geometric graph. Each $u\in 𝒰$ is drawn i.i.d in ${ℛ}^d$ (say a Gaussian centered at 0) and similarly for $v\in 𝒱$. 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 Erdös-Renyi 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 side-information in sequential learning

We end with an open direction that may be relevant to many of the problems considered above: how to use side-information to speed-up 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.

• One-sided 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 127, 61. We will investigate how to incorporate this extra information in RL procedures devised in Theme 1. One option is by leveraging the Kaplan-Meier estimator.
• Side-information in dynamic resource allocation problems and matching
  Generalizing the idea above, one can observe side-information in many other problems 42. 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.

Auctions 96

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 112), 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 127 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 113. 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 60, 111, 64, 86.

Matching 95, 110

A matching is nothing more than a bi-partite 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 vice-versa) and be such that no couple (agent-resource) that are un-paired 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 use-cases

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 use-case. 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.

4.2.1 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 use-cases, in particular the work on advertiser-centric 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 fairness-vs-privacy 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 well-known that usual fairness metrics are not compatible 94. 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 111, 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 e-commerce 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.

6.1.1 Auctions

Most of the display advertising inventory is sold through real-time auctions. The participants of these auctions are typically bidders (Google, Criteo, RTB House, and Trade Desk for instance) that participate on behalf of advertisers. In order to estimate the value of each display opportunity, they usually train advanced machine learning algorithms using historical data. In the labeled training set, the inputs are vectors of features representing each display opportunity, and the labels are the generated rewards. In practice, the rewards are given by the advertiser and are tied to whether a particular user converts. Consequently, the rewards are aggregated at the user level and never observed at the display level. A fundamental task that has, to the best of our knowledge, been overlooked is to account for this mismatch and split, or attribute, the rewards at the right granularity level before training a learning algorithm. We call this the label attribution problem. In 4, we develop an approach to the label attribution problem, which is both theoretically justified and practical. In particular, we develop a fixed point algorithm that allows for large-scale implementation and showcase our solution using a large-scale publicly available data set from Criteo, a large demand-side platform. We dub our approach the fixed point label attribution algorithm. Managerial implications: There is often a hidden leap of faith when transforming the advertiser’s signal into display labeling. Demand Side Platforms providers should be careful when building their machine learning pipeline and carefully solve the label attribution step.

In 8, we use a fixed point gradient flow algorithm to compute the equilibria of first-price procurement auctions in the presence of losses and Bayesian priors. We use this efficient algorithm to compare optimal, first-price and VCG auctions. This allows us to numerically estimate the social cost of sub-optimality of the nodal pricing mechanism in wholesale electricity markets. We also derive a closed form expression of the optimal mechanism procurement cost when the types are uniformly distributed. In 9, we discuss a procurement problem with transportation losses and piecewise linear production costs. We first provide an algorithm based on Knaster-Tarski’s fixed point theorem to solve the allocation problem in the quadratic losses case. We then identify a monotony condition on the types distribution under which the Bayesian cost minimizing mechanism takes a simple form.

In 11, motivated by online display advertising, we consider repeated second-price auctions, where agents sample their value from an unknown distribution with cumulative distribution function F . In each auction t, a decision-maker bound by limited observations selects n t agents from a coalition of N to compete for a prize with p other agents, aiming to maximize the cumulative reward of the coalition across all auctions. The problem is framed as an N -armed structured bandit, each number of player sent being an arm n, with expected reward r(n) fully characterized by F and p + n. We present two algorithms, Local-Greedy (LG) and Greedy-Grid (GG), both achieving constant problem-dependent regret. This relies on three key ingredients: 1. an estimator of r(n) from feedback collected from any arm k, 2. concentration bounds of these estimates for k within an estimation neighborhood of n and 3. the unimodality property of r under standard assumptions on F . Additionally, GG exhibits problem-independent guarantees on top of best problem-dependent guarantees. However, by avoiding to rely on confidence intervals, LG practically outperforms GG, as well as standard unimodal bandit algorithms such as OSUB or multi-armed bandit algorithms.

In 25, motivated by the strategic participation of electricity producers in electricity day-ahead market, we study the problem of online learning in repeated multi-unit uniform price auctions focusing on the adversarial opposing bid setting. The main contribution of this paper is the introduction of a new modeling of the bid space. Indeed, we prove that a learning algorithm leveraging the structure of this problem achieves a regret of $\tilde{O}\left(K^{4/3}{T}^{2/3}\right)$ under bandit feedback, improving over the bound of $\tilde{O}\left(K^{7/4}{T}^{3/4}\right)$ previously obtained in the literature. This improved regret rate is tight up to logarithmic terms. Inspired by electricity reserve markets, we further introduce a different feedback model under which all winning bids are revealed. This feedback interpolates between the full-information and bandit scenarios depending on the auctions' results. We prove that, under this feedback, the algorithm that we propose achieves regret $\tilde{O}\left(K^{5/2}\sqrt{T}\right)$.

6.1.2 Dynamic pricing

In 26, we proposed improved algorithms for contextual dynamic pricing. In contextual dynamic pricing, a seller sequentially prices goods based on contextual information. Buyers will purchase products only if the prices are below their valuations. The goal of the seller is to design a pricing strategy that collects as much revenue as possible. We focus on two different valuation models. The first assumes that valuations linearly depend on the context and are further distorted by noise. Under minor regularity assumptions, our algorithm achieves an optimal regret bound of $\widetilde{𝒪}\left(T^{2/3}\right)$, improving the existing results. The second model removes the linearity assumption, requiring only that the expected buyer valuation is $\beta$-Hölder in the context. For this model, our algorithm obtains a regret $\widetilde{𝒪}\left(T^{d+2\beta /d+3\beta }\right)$, where $d$ is the dimension of the context space.

6.1.3 Mechanism design

In 17, we study auction design within the widely acclaimed model of interdependent values, introduced by Milgrom and Weber [1982]. In this model, every bidder $i$ has a private signal $s_i$ for the item for sale, and a public valuation function $v_i(s_1,...,s_n)$ which maps every vector of private signals (of all bidders) into a real value. A recent line of work established the existence of approximately-optimal mechanisms within this framework, even in the more challenging scenario where each bidder's valuation function $v_i$ is also private. This body of work has primarily focused on single-item auctions with two natural classes of valuations: those exhibiting submodularity over signals (SOS) and $d$-critical valuations. In this work we advance the state of the art on interdependent values with private valuation functions, with respect to both SOS and $d$-critical valuations. For SOS valuations, we devise a new mechanism that gives an improved approximation bound of 5 for single-item auctions. This mechanism employs a novel variant of an “eating mechanism”, leveraging LP-duality to achieve feasibility with reduced welfare loss. For $d$-critical valuations, we broaden the scope of existing results beyond single-item auctions, introducing a mechanism that gives a $(d+1)$-approximation for any environment with matroid feasibility constraints on the set of agents that can be simultaneously served. Notably, this approximation bound is tight, even with respect to single-item auctions.

In 19, we study best-of-both-worlds guarantees for the fair division of indivisible items among agents with subadditive valuations. Our main result establishes the existence of a random allocation that is simultaneously ex-ante $\frac{1}{2}$-envy-free, ex-post $\frac{1}{2}$-EFX and ex-post EF1, for every instance with subadditive valuations. We achieve this result by a novel polynomial-time algorithm that randomizes the well-established envy cycles procedure in a way that provides ex-ante fairness. Notably, this is the first best-of-both-worlds fairness guarantee for subadditive valuations, even when considering only EF1 without EFX.

In 6, we propose and theoretically analyze a measure to encourage greater voluntary contributions to public goods. Our measure is a simple intervention that restricts individuals’ strategy sets by imposing a minimum individual contribution level while still allowing for full free riding for those who do not want to contribute. We show that for a well-chosen value of the minimum individual contribution level, this measure does not incentivize any additional free riding while strictly increasing the total contributions relative to the situation without the minimum contribution level. Our measure is appealing because it is nonintrusive and in line with the principle of “freedom of choice.” It is easily implementable for many different public goods settings where more intrusive measures are less accepted. This approach has been implemented in practice in some applications, such as charities.

6.3.1 Data valuation

In 20, we consider the dataset valuation problem, that is the problem of quantifying the incremental gain, to some relevant pre-defined utility of a machine learning task, of aggregating an individual dataset to others.The Shapley value is a natural tool to perform dataset valuation due to its formal axiomatic justification, which can be combined with Monte Carlo integration to overcome the computational tractability challenges. Such generic approximation methods, however, remain expensive in some cases. In this paper, we exploit the knowledge about the structure of the dataset valuation problem to devise more efficient Shapley value estimators. We propose a novel approximation, referred to as discrete uniform Shapley, which is expressed as an expectation under a discrete uniform distribution with support of reasonable size. We justify the relevancy of the proposed framework via asymptotic and non-asymptotic theoretical guarantees and illustrate its benefits via an extensive set of numerical experiments.

6.3.2 Bandits and reinforcement learning

In 5, we provide a survey of recent results in multi-player bandits.

In 12, we consider the classical multi-armed bandit problem, but with strategic arms. In this context, each arm is characterized by a bounded support reward distribution and strategically aims to maximize its own utility by potentially retaining a portion of its reward, and disclosing only a fraction of it to the learning agent. This scenario unfolds as a game over $T$ rounds, leading to a competition of objectives between the learning agent, aiming to minimize their regret, and the arms, motivated by the desire to maximize their individual utilities. To address these dynamics, we introduce a new mechanism that establishes an equilibrium wherein each arm behaves truthfully and discloses as much of its rewards as possible. With this mechanism, the agent can attain the second-highest average (true) reward among arms, with a cumulative regret bounded by $O(log(T)/\Delta )$ (problem-dependent) or $O\left(\sqrt{Tlog\left(T\right)}\right)$ (worst-case).

In 23, we look at The Value of Reward Lookahead in Reinforcement Learning. In reinforcement learning (RL), agents sequentially interact with changing environments while aiming to maximize the obtained rewards. Usually, rewards are observed only after acting, and so the goal is to maximize the expected cumulative reward. Yet, in many practical settings, reward information is observed in advance – prices are observed before performing transactions; nearby traffic information is partially known; and goals are oftentimes given to agents prior to the interaction. In this work, we aim to quantifiably analyze the value of such future reward information through the lens of competitive analysis. In particular, we measure the ratio between the value of standard RL agents and that of agents with partial future-reward lookahead. We characterize the worst-case reward distribution and derive exact ratios for the worst-case reward expectations. Surprisingly, the resulting ratios relate to known quantities in offline RL and reward-free exploration. We further provide tight bounds for the ratio given the worst-case dynamics. Our results cover the full spectrum between observing the immediate rewards before acting to observing all the rewards before the interaction starts. In 24, we study reinforcement learning (RL) problems in which agents observe the reward or transition realizations at their current state before deciding which action to take. Such observations are available in many applications, including transactions, navigation and more. When the environment is known, previous work shows that this lookahead information can drastically increase the collected reward. However, outside of specific applications, existing approaches for interacting with unknown environments are not well-adapted to these observations. In this work, we close this gap and design provably-efficient learning algorithms able to incorporate lookahead information. To achieve this, we perform planning using the empirical distribution of the reward and transition observations, in contrast to vanilla approaches that only rely on estimated expectations. We prove that our algorithms achieve tight regret versus a baseline that also has access to lookahead information - linearly increasing the amount of collected reward compared to agents that cannot handle lookahead information.

6.4.1 Prophet inequalities

In 13, we study Lookback Prophet Inequalities. Prophet inequalities are fundamental optimal stopping problems, where a decision-maker observes sequentially items with values sampled independently from known distributions, and must decide at each new observation to either stop and gain the current value or reject it irrevocably and move to the next step. This model is often too pessimistic and does not adequately represent real-world online selection processes. Potentially, rejected items can be revisited and a fraction of their value can be recovered. To analyze this problem, we consider general decay functions $D_1,D_2,...$, quantifying the value to be recovered from a rejected item, depending on how far it has been observed in the past. We analyze how lookback improves, or not, the competitive ratio in prophet inequalities in different order models. We show that, under mild monotonicity assumptions on the decay functions, the problem can be reduced to the case where all the decay functions are equal to the same function $x\mapsto \gamma x$, where $\gamma ={inf}_{x>0}{inf}_{j\ge 1}{D}_j\left(x\right)/x$. Consequently, we focus on this setting and refine the analyses of the competitive ratios, with upper and lower bounds expressed as increasing functions of $\gamma$.

In 14, we consider a hiring process with candidates coming from different universities. It is easy to order candidates with the same background, yet it can be challenging to compare them otherwise. The latter case requires additional costly assessments, leading to a potentially high total cost for the hiring organization. Given an assigned budget, what would be an optimal strategy to select the most qualified candidate? We model the above problem as a multicolor secretary problem, allowing comparisons between candidates from distinct groups at a fixed cost. Our study explores how the allocated budget enhances the success probability in such settings.

In 22, we study online selection problems in both the prophet and secretary settings, when arriving agents have interdependent values. In the interdependent values model, introduced in the seminal work of Milgrom and Weber [1982], each agent has a private signal and the value of an agent is a function of the signals held by all agents. Results in online selection crucially rely on some degree of independence of values, which is conceptually at odds with the interdependent values model. For prophet and secretary models under the standard independent values assumption, prior works provide constant factor approximations to the welfare. On the other hand, when agents have interdependent values, prior works in Economics and Computer Science provide truthful mechanisms that obtain optimal and approximately optimal welfare under certain assumptions on the valuation functions. We bring together these two important lines of work and provide the first constant factor approximations for prophet and secretary problems with interdependent values. We consider both the algorithmic setting, where agents are non-strategic (but have interdependent values), and the mechanism design setting with strategic agents. All our results are constructive and use simple stopping rules.

6.4.2 Algorithms with prediction and scheduling

In 15, we address the problem of learning-augmented priority queueus. Priority queues are one of the most fundamental and widely used data structures in computer science. Their primary objective is to efficiently support the insertion of new elements with assigned priorities and the extraction of the highest priority element. In this study, we investigate the design of priority queues within the learning-augmented framework, where algorithms use potentially inaccurate predictions to enhance their worst-case performance. We examine three prediction models spanning different use cases, and we show how the predictions can be leveraged to enhance the performance of priority queue operations. Moreover, we demonstrate the optimality of our solution and discuss some possible applications.

In 16, we consider the non-clairvoyant scheduling problem. The non-clairvoyant scheduling problem has gained new interest within learning-augmented algorithms, where the decision-maker is equipped with predictions without any quality guarantees. In practical settings, access to predictions may be reduced to specific instances, due to cost or data limitations. Our investigation focuses on scenarios where predictions for only $B$ job sizes out of $n$ are available to the algorithm. We first establish near-optimal lower bounds and algorithms in the case of perfect predictions. Subsequently, we present a learning-augmented algorithm satisfying the robustness, consistency, and smoothness criteria, and revealing a novel tradeoff between consistency and smoothness inherent in the scenario with a restricted number of predictions.

In 21, we consider the question: How should an expert send forecasts to maximize her utility subject to passing a calibration test? We consider a dynamic game where an expert sends probabilistic forecasts to a decision maker. The decision maker uses a calibration test based on past outcomes to verify the expert's forecasts. We characterize the optimal forecasting strategy by reducing the dynamic game to a static persuasion problem. A distribution of forecasts is implementable by a calibrated strategy if and only if it is a mean-preserving contraction of the distribution of conditionals (honest forecasts). We characterize the value of information by comparing what an informed and uninformed expert can attain. Moreover, we consider a decision maker who uses regret minimization, instead of the calibration test, to take actions. We show that the expert can achieve the same payoff against a regret minimizer as under the calibration test, and in some instances, she can achieve strictly more.

In 10, we consider algorithms for scheduling deadline-sensitive malleable tasks. Due to the ubiquity of batch data processing, the related problems of scheduling malleable batch tasks have received significant attention. We consider a fundamental model where a set of tasks is to be processed on multiple identical machines and each task is specified by a value, a workload, a deadline and a parallelism bound. Within the parallelism bound, the number of machines assigned to a task can vary over time without affecting its workload. In this paper, we identify a boundary condition and prove by construction that a set of malleable tasks with deadlines can be finished by their deadlines if and only if it satisfies the boundary condition. This core result plays a key role in the design and analysis of scheduling algorithms: (i) when several typical objectives are considered, such as social welfare maximization, machine minimization, and minimizing the maximum weighted completion time, and, (ii) when the algorithmic design techniques such as greedy and dynamic programming are applied to the social welfare maximization problem. As a result, we give four new or improved algorithms for the above problems.

7.1.1 Visits of international scientists

7.1.2 Visits to international teams

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:
  2023-2027 (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:
  2021-2025 (4 years)
• Additionnal info/keywords:
  ANR JCJC project, 245k euros. Fairness, matching, auctions.

BOLD (ANR)

Participants: Vianney Perchet.

• Title:
  BOLD: Beyond Online Learning for Better Decisions
• Partner Institution(s):
  • Crest, Genes
• Date/Duration:
  2019-2024 (4.5 years)
• Additionnal info/keywords:
  ANR project, 270k euros. online learning, optimization, bandits.

DOOM (ANR)

Participants: Vianney Perchet.

• Title:
  DOOM: Design of Optimal Online Matching Markets
• Partner Institution(s):
  • Crest, Genes
• Date/Duration:
  2014-2028 (4 years)
• Additionnal info/keywords:
  ANR project, 456k euros. online markets, learning, matching.

8.1.1 Scientific events: organisation

8.1.2 Scientific events: selection

8.1.3 Journal

8.1.4 Invited talks

• Cristina Butucea
  Anniversary Conference V. Koltchinskii (Bertinoro, Italy), ISNP (Braga, Portugal), MFOberwolfach Germany, Journées de la SMAI (Poitiers), University of Nanjing (China)
• Hugo Richard
  PGMO Days (France)
• Simon Mauras
  Junior Workshop on Mathematical Game Theory (Rome), Algorithms, Learning, and Games (Sicily), Alpine Game Theory Symposium (Grenoble), Game theory seminar (Paris), Optimization seminar (Paris)

8.1.5 Scientific expertise

• Vianney Perchet:
  Expert for the evaluation of the LABEX MME:DII
• Patrick Loiseau:
  External expert for the ANR and the Royal Society
• Cristina Butucea:
  ANR evaluation committee, hiring committees France and Germany,

8.2.1 Supervision

• Patrick Loiseau:
  PhD students: Rémi Castera, Mathieu Molina, Reda Jalal, Minrui Xu, Marie Generali, Melissa Tamine; postdocs: Felipe Garrido Lucero, Simon Finster, Denis Sokolov
• Vianney Perchet:
  PhD students: Come Fiegel, Maria Cherifa, Mathieu Molina, Ziyad Benomar, Mike Liu, Hafedh El Ferchichi, Matilde Tullii, Giovanni Montanari. postdocs: Felipe Garrido Lucero, Nadav Merlis, Solenne Gaucher, Dorian Baudry
• Matthieu Lerasle
  PhD Students: Clara Carlier, Hugo Chardon, Hafedh El Ferchichi.
• Cristina Butucea
  PhD students: Nayel Bettache, Henning Stein, Antoine Schoonaert
• Marc Abeille:
  Ahmed Ben Yahmed
• Clément Calauzènes:
  Morgane Goibert, Maria Cherifa
• Benjamin Heymann:
  Mélissa Tamine
• Maxime Vono:
  Mélissa Tamine
• Simon Mauras:
  PhD student: Minrui Xu.

8.4.1 Specific official responsibilities in science outreach structures

• Simon Mauras:
  • Future co-responsable médiation at INRIA Saclay.
  • Volunteer and treasurer of association France-IOI: training and selecting the french team for the International Olympiads in Informatics; organizing summer camps to teach algorithmic to groups of $\sim$20 highschool students, creating online contests and educational content about informatics (Concours castor, concours Alkindi, Concours Algoréa, online content for SNT option in highschool)

International journals

• 4 articleM.Martin Bompaire, A.Antoine Désir and B.Benjamin Heymann. Fixed Point Label Attribution for Real-Time Bidding.Manufacturing and Service Operations Management263May 2024, 1043-1061HALDOIback to text
• 5 articleE.Etienne Boursier and V.Vianney Perchet. A survey on multi-player bandits.Journal of Machine Learning ResearchJanuary 2024HALback to text
• 6 articleM.Michela Chessa and P.Patrick Loiseau. Enhancing voluntary contributions in a public goods economy via a minimum individual contribution level.Public Choice2011-2April 2024, 237-261HALDOIback to text
• 7 articleC.Christophe Giraud, Y.Yann Issartel, L.Luc Lehericy and M.Matthieu Lerasle. Pair-Matching: Link Prediction with Adaptive Queries.Mathematical Statistics and LearningMarch 2024HALback to text
• 8 articleB.Benjamin Heymann and A.Alejandro Jofré. Procurement auctions with losses.Computational Management Science212July 2024, 38HALDOIback to text
• 9 articleB.Benjamin Heymann and A.Alejandro Jofré. Reverse auctions with transportation and convex costs.Computational Management Science211February 2024, 20HALDOIback to text
• 10 articleX.Xiaohu Wu and P.Patrick Loiseau. Algorithms for Scheduling Deadline-Sensitive Malleable Tasks.SN Operations Research Forum52April 2024, 30HALDOIback to text

International peer-reviewed conferences

• 11 inproceedingsD.Dorian Baudry, H.Hugo Richard, M.Maria Cherifa, C.Clément Calauzènes and V.Vianney Perchet. Optimizing the coalition gain in Online Auctions with Greedy Structured Bandits.NeurIPS 2024NeurIPS 2024 - 38th Conference on Neural Information Processing SystemsVancouver (BC), CanadaDecember 2024HALback to text
• 12 inproceedingsA.Ahmed Ben Yahmed, C.Clément Calauzènes and V.Vianney Perchet. Strategic Multi-Armed Bandit Problems Under Debt-Free Reporting.NeurIPS 2024 - 38th Annual Conference on Neural Information Processing SystemsVancouver, CanadaDecember 2024HALback to text
• 13 inproceedingsZ.Ziyad Benomar, D.Dorian Baudry and V.Vianney Perchet. Lookback Prophet Inequalities.NeurIPS 2024 - 38th Conference on Neural Information Processing SystemsVancouver (CA), CanadaDecember 2024HALback to text
• 14 inproceedingsZ.Ziyad Benomar, E.Evgenii Chzhen, N.Nicolas Schreuder and V.Vianney Perchet. Addressing Bias in Online Selection with Limited Budget of Comparisons.PMLRNeurIPS 2024 - 38th Conference on Neural Information Processing SystemsVancouver (CA), CanadaDecember 2024HALback to text
• 15 inproceedingsZ.Ziyad Benomar and C.Christian Coester. Learning-Augmented Priority Queues.NeurIPS 2024 - 38th Conference on Neural Information Processing SystemsVancouver (CA), CanadaDecember 2024HALback to text
• 16 inproceedingsZ.Ziyad Benomar and V.Vianney Perchet. Non-clairvoyant Scheduling with Partial Predictions.Proceedings of the 41 st International Conference on MachineLearning, Vienna, Austria. PMLR 235, 2024ICML 2024 - The 41st International Conference on Machine LearningVienna, AustriaJuly 2024HALback to text
• 17 inproceedingsA.Alon Eden, M.Michal Feldman, S.Simon Mauras and D.Divyarthi Mohan. Private Interdependent Valuations: New Bounds for Single-Item Auctions and Matroids.EC '24: Proceedings of the 25th ACM Conference on Economics and ComputationEC 2024 - 25th ACM Conference on Economics and ComputationNew Haven, United StatesACMJuly 2024, 448-464HALDOIback to text
• 18 inproceedingsH.Hafedh El Ferchichi, M.Matthieu Lerasle and V.Vianney Perchet. Active Ranking and Matchmaking, with Perfect Matchings.Proceedings of the 41 st International Conference on Machine Learning, Vienna, Austria. PMLR 235, 2024ICML 2024 - The Forty-First International Conference on Machine LearningVienne, AustriaMay 2024HALback to text
• 19 inproceedingsM.Michal Feldman, S.Simon Mauras, V. V.Vishnu V Narayan and T.Tomasz Ponitka. Breaking the Envy Cycle: Best-of-Both-Worlds Guarantees for Subadditive Valuations.EC '24: Proceedings of the 25th ACM Conference on Economics and ComputationEC 2024 - 25th ACM Conference on Economics and ComputationNew Haven, United StatesACMJuly 2024, 1236 - 1266HALDOIback to text
• 20 inproceedingsF.Felipe Garrido-Lucero, B.Benjamin Heymann, M.Maxime Vono, P.Patrick Loiseau and V.Vianney Perchet. DU-Shapley: A Shapley Value Proxy for Efficient Dataset Valuation.Advances in Neural Information Processing Systems 37 (NeurIPS 2024)Advances in Neural Information Processing Systems 37 (NeurIPS 2024)Vancouver, Canada2024HALback to text
• 21 inproceedingsA.Atulya Jain and V.Vianney Perchet. Calibrated Forecasting and Persuasion.EC 2024 - The Twenty-Fifth ACM Conference on Economics and ComputationNew haven, United StatesJune 2024HALback to text
• 22 inproceedingsS.Simon Mauras, D.Divyarthi Mohan and R.Rebecca Reiffenhäuser. Optimal Stopping with Interdependent Values.EC '24: Proceedings of the 25th ACM Conference on Economics and ComputationEC 2024 - 25th ACM Conference on Economics and ComputationNew Haven, United StatesACMJuly 2024, 246-265HALDOIback to text
• 23 inproceedingsN.Nadav Merlis, D.Dorian Baudry and V.Vianney Perchet. The Value of Reward Lookahead in Reinforcement Learning.NeurIPS 2024 - 38th Conference on Neural Information Processing SystemsVancouver, CanadaDecember 2024HALback to text
• 24 inproceedingsN.Nadav Merlis. Reinforcement Learning with Lookahead Information.NeurIPS 2024 - 38th Conference on Neural Information Processing SystemsVancouver, CanadaDecember 2024HALback to text
• 25 inproceedingsM.Marius Potfer, D.Dorian Baudry, H.Hugo Richard, V.Vianney Perchet and C.Cheng Wang. Improved learning rates in multi-unit uniform price auctions.NeurIPS 2024NeurIPS 2024 - 38th Conference on Neural Information Processing SystemsVancouver (BC), CanadaDecember 2024HALback to text
• 26 inproceedingsM.Matilde Tullii, S.Solenne Gaucher, N.Nadav Merlis and V.Vianney Perchet. Improved Algorithms for Contextual Dynamic Pricing.NeurIPS 2024 - 38th Conference on Neural Information Processing SystemsNeurIPS 2024 - The 38th Annual Conference on Neural Information Processing SystemsVancouver (BC), CanadaarXiv2024HALback to text

Doctoral dissertations and habilitation theses

• 27 thesisR.Rémi Castera. Efficiency and Fairness in Matching Problems.UGA (Université Grenoble Alpes)October 2024HAL

Reports & preprints

• 28 miscC.Charles Arnal, V.Vivien Cabannes and V.Vianney Perchet. Mode Estimation with Partial Feedback.2024HALDOI
• 29 miscD.Dorian Baudry, N.Nadav Merlis, M.Mathieu Molina, H.Hugo Richard and V.Vianney Perchet. Multi-Armed Bandits with Guaranteed Revenue per Arm.January 2024HAL
• 30 miscN.Nayel Bettache and C.Cristina Butucea. Two-sided Matrix Regression.January 2024HAL
• 31 miscC.Cristina Butucea, J.-F.Jean-François Delmas, A.Anne Dutfoy and C.Clément Hardy. Off-the-grid prediction and testing for linear combination of translated features.July 2024HAL
• 32 miscC.Cristina Butucea, J.-F.Jean-François Delmas, A.Anne Dutfoy and C.Clément Hardy. Simultaneous off-the-grid learning of mixtures issued from a continuous dictionary.February 2024HAL
• 33 miscR.Rémi Castera, P.Patrick Loiseau and B.Bary Pradelski. Correlation of Rankings in Matching Markets.February 2024HAL
• 34 miscV.Vincent Divol and S.Solenne Gaucher. Demographic parity in regression and classification within the unawareness framework.September 2024HAL
• 35 miscS.Solenne Gaucher, G.Gilles Blanchard and F.Frédéric Chazal. Supervised Contamination Detection, with Flow Cytometry Application.April 2024HAL
• 36 miscB.Benjamin Heymann, R.Rémi Chan--Renous-Legoubin and A.Alexandre Gilotte. A pragmatic policy learning approach to account for users' fatigue in repeated auctions.July 2024HAL
• 37 miscB.Benjamin Heymann and M.Marc Lanctot. Learning in Games with progressive hiding.September 2024HAL
• 38 miscM.Matilde Tullii, S.Solenne Gaucher, H.Hugo Richard, E.Eustache Diemert, V.Vianney Perchet, A.Alain Rakotomamonjy, C.Clément Calauzènes and M.Maxime Vono. Position Paper: Open Research Challenges for Private Advertising Systems under Local Differential Privacy.February 2024HAL