How can we ensure that an advertisement or a job offer can be seen in a non-discriminatory way by the people they are displayed to?
Inria, ENSAE (National School of Statistics and Economics, part of the Institut Polytechnique de Paris), and the Adtech leader Criteo founded the FAIRPLAY program in 2022. This groundbreaking initiative addresses pivotal challenges in machine learning: developing systems that learn efficiently while ensuring fairness and privacy in multi-agent environments, whether agents are competing, cooperating, or both (“coopetiting”).
How do you avoid discrimination while protecting user data?
The FAIRPLAY project-team was created as part of a research partnership between the Criteo AI Lab and Inria. Its…
www.inria.fr
Efficiency, fairness, and privacy are the cornerstones of Criteo’s marketplace solutions, including CGrowth, CMax, CGrid, and CYield, all of which rely heavily on advanced machine learning models. FAIRPLAY dedicated to take into account these principles into the core of machine learning systems, responding to the growing need for ethical and high-performing algorithms in diverse applications, from online recommendations to advertising auctions.
This collaboration seeks to design machine learning algorithms that respect both users and the data they provide. With the increasing prevalence of machine learning, it is imperative that these systems operate not only efficiently but also ethically. FAIRPLAY’s goal is to integrate mechanisms that enhance fairness and safety into the learning processes, achieving these aims at minimal cost. The joint research efforts aim to reconcile the often conflicting demands of fairness and performance.
FAIRPLAY frontally addresses the complexity of marketplaces that live by interactions between agents.. These agents, who are simultaneously optimizing their own objectives, can exhibit both cooperative and competitive behaviors, which may lead to the depreciation or manipulation of learning algorithms. The program emphasizes the development of systems capable of functioning efficiently without centralized access to complete information, reflecting the decentralized nature of many decision-making processes.
Addressing the ethical constraints in multi-agent learning systems requires a multidisciplinary approach. FAIRPLAY integrates and advances theoretical tools and techniques from game theory, economics, online and deep reinforcement learning, optimal transport, competitive analysis, and theoretical computer science. This comprehensive strategy ensures the development of machine learning systems that are not only powerful and effective but also fair and respectful of privacy.
A Few Examples of Research Projects Explored
Data Cooperation: A Shapley Value Proxy for Efficient Dataset Valuation
In the field of digital advertising, learning systems rely on data provided by advertisers, users, and markets. The question is how to encourage these actors, regardless of the volume of their data (low or high), to share their information in exchange for a greater benefit compared with what they would get by keeping their data to themselves. Indeed, the amount of information, i.e., the number of datapoint is crucial to take more accurate and efficient decisions. By grouping their dataset, the different actors will learn, in a centralized fashion, a better decision mechanism that will, once implemented generate higher revenue to all of them. This said, some actors — either because they are larger or because their data are of better quality — might benefit less than others (the smaller ones).
The solution we are exploring is to reward each actor based on their individual contribution to the efficiency of the learning system. There are concepts of cooperative game theory to theoretically measure and quantify the money exchanges to be implemented (how much the small actors should give to the larger ones to “recompensate”, and incentivize, them to participate). The major downside is that the computation cost of those solutions is highly prohibitive, thus they are not tractable. We devised efficient approximations of them that can easily be implementable.
A byproduct of our methods is that actors can choose “how much” data they actually provide, by carefully adding noising their dataset, using differential privacy techniques.
This project spawned a research paper that has just been accepted at NeurIPS’2024 !
DU-Shapley: A Shapley Value Proxy for Efficient Dataset Valuation
We consider the dataset valuation problem, that is, the problem of quantifying the incremental gain, to some relevant…
arxiv.org
Private advertising systems under Local Differential Privacy
In the absence of third-party cookies, the online advertisers (including Criteo) will rely on a variety of sources, including a set of APIs, to access the data. This leads to work with aggregated user-level data, which includes past interactions such as clicks, rather than granular user information. In addition, machine learning model training is outsourced to trusted servers.
We are addressing the issue of noisy event-level reporting by applying local Differential Privacy to features and labels. This framework introduces noise to the features, requiring minimal engineering changes as the pipelines remain consistent. This approach can be applied to both discrete and continuous information, offering several types of mechanisms.
The introduction of noise is a one-time process, not a recurring computation, making it compatible with our current tech stack. One of our objectives is to train multiple machine learning models using the noisy data we obtain from users. This presents a trade-off between the information we want to query and its reuse for multiple tasks.
Another challenge is working with a fixed dataset while aiming to perform a variety of tasks. The first step is to explore the data to make informed queries. For instance, in order to predict the price of a flat, you might need to select only three relevant features from a query containing dozens of features. This requires a strategy of exploration and exploitation to design your query based on prior data and knowledge.
We explain some of those questions in a research paper published at WISE’2024 !
Position Paper: Open Research Challenges for Private Advertising Systems under Local Differential…
Due to the ongoing deprecation of third-party cookies on mainstream browsers, the digital advertising industry is…
polytechnique.hal.science
Trading-off price for data quality to achieve fair online allocation
User data can be quite sensitive for two different, opposite reasons. The first one is that they actually generate and possess those data and therefore have the right to decide how they should be used, and more importantly, which part of it could, or not, be used. This is the concept of “privacy”: a user should be able to decide whether some sensitive attribute (the gender, ethnicity, or any other features) can be revealed and/or be part of some learning algorithm. The second sensitive property of user data is that each user should be treated similarly (assuming their relevant features are the same); this is the concept of “fairness”. Of course, both concepts are desirable (one on the individual level, the other one on a global level) but they are clearly antagonist: privacy requires access to sensitive irrelevant feature (say, the gender) while privacy aims at protecting them.
We investigated how, and at what cost, they could be reconciled. A tentative solution is quite natural: users are willing to share some private information, if they are rewarded for this. As a consequence, a learning algorithm should balance the cost of acquiring date with the gain of being fair.
Our results were published in this research paper at the conference NeurIPS’2023.
Trading-off price for data quality to achieve fair online allocation
Trading-off price for data quality to achieve fair online allocation Part of Advances in Neural Information Processing…
proceedings.neurips.cc
What’s next?
The FAIRPLAY team is also investigating many other different projects related to competitive systems. Among others, we can mention the following 👇
Online Private Auctions
In the near future, when the Privacy SandBox of Chrome will be put in place, it will be quite often impossible to track and identify the user browsing online. This will have a huge impact on the mechanism of auctions used to determine who can display an ad (Criteo, Google, etc.) on some given website. The DSP, such as Criteo, will provide “bidding strategies” instead of directly sending bids, changing drastically the strategy space. In this paper, we investigate how the optimal bidding strategies can be learned. A first paper on this research direction has been accepted at NeurIPS’2024 !
Learning in repeated auctions
Auction theory historically focused on the question of designing the best way to sell a single item to potential…
hal.science
Fair Division of Revenue in Repeated Ressource Allocation
Recommender systems can be viewed as a resource (the users) allocation (to the different content/items) problem. When repeated, finding the “optimal” content to recommend can be recast as a bandit problem. The main issue is that the recommender systems tend to be polarized towards the “most promising content”, i.e., those that are the most viewed/clicked for instance. Unfortunately, a naïve implementation of such algorithms will end up in a system with a few diversities (as the smaller, less popular contents tend to disappear). In this project, we aim at guaranteeing, when possible, some fixed revenue to each content, to maintain diversity. This started with a paper at AIStats’2024.
Multi-Armed Bandits with Guaranteed Revenue per Arm
We consider a Multi-Armed Bandit problem with covering constraints, where the primary goal is to ensure that each arm…
theses.hal.science
Active Labelling with Annotators
A typical problem in industry is that the data generating processes evolves with time, for instance, new products appear or disappear, trends die, etc. As a consequence, datasets that used to be relevant for a learning problem deteriorate with time and eventually end up with a distribution quite different from the current distribution of data. It is therefore required to label new data-points and/or to use transfer learning (from the old to the new data distributions). These labelings have a cost, as they are quite often human-made. We investigate how to build the best possible dataset, at the smallest cost possible (for instance, by avoiding labeling new data points on which we are already quite certain of their true label, or those that provide less information). This is even more challenging when the set of labels is large, and has some hierarchical structure, as in the Criteo use case. Some of our findings are summarized in a paper published at COLT’2024 and NeurIPS’2023.
Active Labeling: Streaming Stochastic Gradients
Active Labeling: Streaming Stochastic Gradients Part of Advances in Neural Information Processing Systems 35 (NeurIPS…
proceedings.neurips.cc
Mode Estimation with Partial Feedback
The combination of lightly supervised pre-training and online fine-tuning has played a key role in recent AI…
arxiv.org
