This year, our research team got a record of 15 papers accepted at the major ML conference.
The 38th annual edition of the prestigious international conference NeurIPS 2024 (Neural Information Processing Systems) took place in Vancouver. ๐ฅ๐ฒ๐๐ฒ๐ฎ๐ฟ๐ฐ๐ต๐ฒ๐ฟ๐ ๐ณ๐ฟ๐ผ๐บ ๐๐ต๐ฒ ๐๐ฟ๐ถ๐๐ฒ๐ผ ๐๐ ๐๐ฎ๐ฏ were there to present their latest ML papers (๐ญ๐ฏ ๐ฝ๐ฎ๐ฝ๐ฒ๐ฟ๐โโโ๐ฎ๐บ๐ผ๐ป๐ด ๐๐ต๐ถ๐ฐ๐ต ๐ผ๐ป๐ฒ ๐๐ฝ๐ผ๐๐น๐ถ๐ด๐ต๐โโโ, ๐ญ ๐ฑ๐ฎ๐๐ฎ๐๐ฒ๐ ๐ฝ๐ฎ๐ฝ๐ฒ๐ฟ, ๐ญ ๐ฐ๐ต๐ฎ๐น๐น๐ฒ๐ป๐ด๐ฒ ๐ฝ๐ฎ๐ฝ๐ฒ๐ฟ).
Initial remarks by Liva Ralaivola (VP Research)
Our papers bring original AI contributions on topics a leading AdTech company like Criteo must have expertise onโโโonline learning, bidding, pricing, dataset valuation, fairness, and deep learning.
Thatโs what we do here at Criteo. We are leading the way for others with an academic center of technical and scientific research. We target ๐ฒ๐ป๐ฑ-๐๐ผ-๐ฒ๐ป๐ฑ ๐๐ (for AdTech) proficiency from ๐๐ฝ๐๐๐ฟ๐ฒ๐ฎ๐บ ๐๐ป๐ฑ๐ฒ๐ฟ๐๐๐ฎ๐ป๐ฑ๐ถ๐ป๐ด & ๐ฐ๐ฟ๐ฎ๐ณ๐๐ถ๐ป๐ด ๐๐ต๐ฟ๐ผ๐๐ด๐ต ๐ฝ๐ฟ๐ผ๐ฑ๐๐ฐ๐ ๐ถ๐ป๐๐ฒ๐ด๐ฟ๐ฎ๐๐ถ๐ผ๐ป ๐๐ผ ๐ฟ๐ผ๐ฏ๐๐๐ ๐ฑ๐ฒ๐ฝ๐น๐ผ๐๐บ๐ฒ๐ป๐.
We do that within a concerted and exciting R&D x Product x Sales effort. Because as H.E. Lucckock put it, โ๐๐ฐ ๐ฐ๐ฏ๐ฆ ๐ค๐ข๐ฏ ๐ธ๐ฉ๐ช๐ด๐ต๐ญ๐ฆ ๐ข ๐ด๐บ๐ฎ๐ฑ๐ฉ๐ฐ๐ฏ๐บ. ๐๐ต ๐ต๐ข๐ฌ๐ฆ๐ด ๐ข๐ฏ ๐ฐ๐ณ๐ค๐ฉ๐ฆ๐ด๐ต๐ณ๐ข ๐ต๐ฐ ๐ฑ๐ญ๐ข๐บ ๐ช๐ต.โ
Criteo contributions
Check out all our contributions, with some of them commented on by their authors.
Logarithmic Smoothing for Pessimistic Off-Policy Evaluation, Selection and Learning. (Spotlight paperย ๐)
We are excited to share our recent work on logarithmic smoothing for pessimistic off-policy evaluation, selection, and learning, which offers a fresh perspective on how to confidently improve policies in interactive systems using historical data. In this study, we move beyond standard point estimators and introduce novel, fully empirical concentration bounds for a broad class of importance-weighting risk estimatorsโโโan approach that not only unifies existing methods but also paves the way for new ones.
Building on these insights, we propose a new estimator (LS) that leverages logarithmic smoothing for large importance weights, and demonstrate how this leads to tighter bounds compared to its counterparts. The result is a more robust framework for policy selection and learning, enabling practitioners to deploy improved strategies with higher confidence in real-world applications.
We believe these findings highlight the importance of adopting a pessimistic perspectiveโโโone that prioritizes worst-case outcomesโโโto guide policy decisions. By sharing our techniques and extensive experimental results, we aim to advance the state of offline contextual bandit research and help foster more reliable and principled data-driven decision making.
More details ๐
Research Card: Logarithmic Smoothing for Pessimistic Off-Policy Evaluation, Selection, and Learning
We discuss a new method for off-policy evaluation, selection, and learning in interactive systems.
techblog.criteo.com
DU-Shapley: A Shapley Value Proxy for Efficient Dataset Valuation
(Felipe Garrido-Lucero, Benjamin Heymann, Maxime Vono, Patrick Loiseau, Vianney Perchet)
We are thrilled to present our work on dataset valuation. Datasets are the lifeblood of AI, and often, merging multiple datasets is crucial for enhancing model performance. As contributors of datasets increase the modelโs value, the question arises: how should the benefits be fairly distributed?
Fortunately, game theory offers a principled solution through the Shapley value. However, applying the Shapley value in this context is computationally intractable. In our work, we simplify the Shapley value calculation to make it feasible for dataset valuation, providing a practical approach to assessing and allocating the value contributed by each dataset.
This work will be key to shape the future of data markets. https://arxiv.org/pdf/2306.02071
Optimizing the coalition gain in Online Auctions with Greedy Structured Bandits
(Dorian Baudry, Hugo Richard, Maria Cherifa, Vianney Perchet, Clรฉment Calauzรจnes)
We are excited to share our work on optimizing the gain of a coalition in repeated second-price auctions. In this problem, we take the point of view of a central agent that decides how many participants (from the same team) should be sent to participate in a second price auction against a known number of competitors. All participants are assumed to bid according to an identical (but unknown) distribution. Intuitively, if many members of the coalition participate, they will compete against each other leading to a small gain when the auction is won. On the other hand, if the coalition sends only a few members, the likelihood of winning the auction is low. The goal is then to optimize the number of members to send. We design new concentration inequalities and propose an algorithm that has constant regret (this is an upper bound on the performance that does not depend on the number of repeated auctions).
Optimizing the coalition gain in Online Auctions with Greedy Structured Bandits
Motivated by online display advertising, this work considers repeated second-price auctions, where agents sample theirโฆ
hal.science
Improved learning rates in multi-unit uniform priceย auctions
(Marius Potfer, Dorian Baudry, Hugo Richard, Vianney Perchet, Cheng Wan)
We are excited to share our work on learning how to bid in repeated multi-unit price auctions. In this type of auction, K identical items are for sale. Buyers simultaneously reveal their bids i.e. the price they are willing to pay to acquire 1, 2,ย โฆ, K items. The seller then sets the price to be equal to the K-th highest bid and gives the items to the highest bidding buyers. In this work, we take the point of view of one buyer and see other buyers as an oblivious adversary. Thanks to a new modeling of the bid space, we designed a new algorithm that reaches better theoretical performance (as measured by the expected regret) than previous approaches. More precisely our performance is optimal in terms of its dependence on the number of auctions (horizon).
Improved learning rates in multi-unit uniform price auctions
Motivated by the strategic participation of electricity producers in electricity day-ahead market, we study the problemโฆ
arxiv.org
FairJob: A Real-World Dataset for Fairness in Onlineย Systems
(Mariia Vladimirova, Federico Pavone, Eustache Diemert)
We are proud to be the first to open-source a fairness-aware dataset for advertising, one of the largest fairness-aware tabular datasets released by a company. In our research, we provided an in-depth analysis of where unfairness can arise throughout the advertising process, offering insights that had not been previously explored with such detail, all based on our internal studies.
Building on these findings, we introduced a novel utility metric that accounts for selection bias and shared reproducible experimental results using the open-sourced dataset.
Our study highlights the potential presence of market bias, which could have far-reaching implications for the entire advertising ecosystem. We believe that our work will contribute to the advancement of responsible advertising and provide greater transparency into its underlying processes.
More details ๐
Research Card: FairJob – A Real-World Dataset for Fairness in Online Systems
Research Card on how to make the outcome prediction in online systems fairer?
techblog.criteo.com
Towards Efficient and Optimal Covariance-Adaptive Algorithms for Combinatorial Semi-Bandits
(Julien Zhou, Pierre Gaillard, Thibaud Rahier, Houssam Zenati, Julyan Arbel)
We are glad to share our work on algorithms for Combinatorial Semi-Bandits. In this work, we explore cases when there is a correlation structure between the rewards of the items to be chosen from.
In particular, we propose two algorithms: OLS-UCB-V and COS-V, designed to leverage this structure. They work by estimating covariance matrices at the same time as they make decisions. These estimations are designed speed-up exploration compared to algorithms not taking them into account. Both algorithms are based on the โoptimismโ principle. While OLS-UCB-V is deterministic, COS-V also uses a sampling strategy, which makes it more computationally efficient.
Combinatorial bandits are commonly used to model online advertisement use cases (any problem with repeated choices among supersets, like choosing a small basket from a lot of items) and understanding well how to design good strategies for such problems is important for Criteo.
Towards Efficient and Optimal Covariance-Adaptive Algorithms for Combinatorial Semi-Bandits
We address the problem of stochastic combinatorial semi-bandits, where a player selects among P actions from the powerโฆ
arxiv.org
Strategic Multi-Armed Bandit Problems Under Debt-Free Reporting
(Ahmed Ben Yahmed, Clรฉment Calauzรจnes, Vianney Perchet)
We are excited to share our work on Strategic Multi-Armed Bandit Problems Under Debt-Free Reporting. In this study, we consider a setting where the arms are strategic, meaning they are rational entities seeking to maximize their own utilities. Consequently, each arm has the ability to withhold a portion of its reward and report only a fraction of it to the player.
This setting induces a game between the player, who aims to minimize regret, and the arms, which seek to maximize their utilities. In such a framework, classic low-regret multi-armed bandit (MAB) algorithms suffer from linear regret, as a detrimental collusive equilibrium may arise. To address this challenge, we propose an algorithm that combines sequential learning with mechanism design by incorporating incentives to encourage truthful reporting. Our algorithm achieves a truthful subgame perfect equilibrium (SPE), under which the player experiences lower regret compared to existing methods in the literature.
This framework has broad real-world applications, particularly in scenarios where strategic behavior is prevalent, such as in auctions, e-commerce, and other multi-agent systems. It also serves as a generalization of the principal-agent theory to multi-agent settings.
Strategic Multi-Armed Bandit Problems Under Debt-Free Reporting
We consider the classical multi-armed bandit problem, but with strategic arms. In this context, each arm isโฆ
arxiv.org
Donโt miss the rest of our contributions ๐
AROMA: Preserving Spatial Structure for Latent PDE Modeling with Local Neural Fields
(Louis Serrano, Thomas X Wang, Etienne Le Naour, Jean-Noรซl Vittaut, Patrick Gallinari)
AROMA: Preserving Spatial Structure for Latent PDE Modeling with Local Neural Fields
We present AROMA (Attentive Reduced Order Model with Attention), a framework designed to enhance the modeling ofโฆ
arxiv.org
Boosting Generalization in Parametric PDE Neural Solvers through Adaptive Conditioning
(Armand Kassaรฏ Koupaรฏ, Jorge Mifsut Benet, Yuan Yin, Jean-Noรซl Vittaut, Patrick Gallinari)
GEPS: Boosting Generalization in Parametric PDE Neural Solvers through Adaptive Conditioning
Solving parametric partial differential equations (PDEs) presents significant challenges for data-driven methods due toโฆ
arxiv.org
NeurIPS 2024 ML4CFD Competition: Harnessing Machine Learning for Computational Fluid Dynamics in Airfoil Design
(Mouadh Yagoubi, David Danan, Milad Leyli-abadi, Jean-Patrick Brunet, Jocelyn Ahmed Mazari, Florent Bonnet, maroua gmati, Asma Farjallah, Paola Cinnella, Patrick Gallinari, Marc Schoenauer)
NeurIPS 2024 ML4CFD Competition: Harnessing Machine Learning for Computational Fluid Dynamics inโฆ
The integration of machine learning (ML) techniques for addressing intricate physics problems is increasinglyโฆ
arxiv.org
Local and Adaptive Mirror Descents in Extensive-Form Games
(Cรดme Fiegel, Pierre Mรฉnard, Tadashi Kozuno, Remi Munos, Vianney Perchet, Michal Valko)
Local and adaptive mirror descents in extensive-form games
We study how to learn $ฮต$-optimal strategies in zero-sum imperfect information games (IIG) with trajectory feedback. Inโฆ
arxiv.org
Lookback Prophet Inequalitites
(Ziyad Benomar, Dorian Baudry, Vianney Perchet)
Lookback Prophet Inequalities
Prophet inequalities are fundamental optimal stopping problems, where a decision-maker observes sequentially items withโฆ
arxiv.org
The Value of Reward Lookahead in Reinforcement Learning
(Nadav Merlis, Dorian Baudry, Vianney Perchet)
The Value of Reward Lookahead in Reinforcement Learning
In reinforcement learning (RL), agents sequentially interact with changing environments while aiming to maximize theโฆ
arxiv.org
Improved Algorithms for Contextual Dynamic Pricing
(Matilde Tullii, Solenne Gaucher, Nadav Merlis, Vianney Perchet)
Improved Algorithms for Contextual Dynamic Pricing
In contextual dynamic pricing, a seller sequentially prices goods based on contextual information. Buyers will purchaseโฆ
arxiv.org
Addressing bias in online selection with limited budget of comparisons
(Ziyad Benomar, Evgenii Chzhen, Nicolas Schreuder, Vianney Perchet)
Addressing Bias in Online Selection with Limited Budget of Comparisons
Consider a hiring process with candidates coming from different universities. It is easy to order candidates with theโฆ
arxiv.org
Check the next video by Vianney Perchet, a Principal Research at Criteo and professor at Ensae, where he shares his insights on all our contributions at NeurIPS 2024.
Do you want to be part of this amazing team? ๐ ๐
Research Intern
What You'll Do: The Criteo AI Lab is pioneering innovations in computational advertising. As the center ofโฆ
careers.criteo.com
