To address unfairness issues in federated learning (FL), contemporary approaches typically use frequent model parameter updates and transmissions between the clients and server. In such a process, client-specific information (e.g., local dataset size or data-related fairness metrics) must be sent to the server to compute, e.g., aggregation weights. All of this results in high transmission costs and the potential leakage of client information. As an alternative, we propose a strategy that promotes fair predictions across clients without the need to pass information between the clients and server iteratively and prevents client data leakage. For each client, we first use their local dataset to obtain a synthetic dataset by solving a bilevel optimization problem that addresses unfairness concerns during the learning process. We then pass each client's synthetic dataset to the server, the collection of which is used to train the server model using conventional machine learning techniques (that do not take fairness metrics into account). Thus, we eliminate the need to handle fairness-specific aggregation weights while preserving client privacy. Our approach requires only a single communication between the clients and the server, thus making it computationally cost-effective, able to maintain privacy, and able to ensuring fairness. We present empirical evidence to demonstrate the advantages of our approach. The results illustrate that our method effectively uses synthetic data as a means to mitigate unfairness and preserve client privacy.

Creating fair AI systems is a complex problem that involves the assessment of context-dependent bias concerns. Existing research and programming libraries express specific concerns as measures of bias that they aim to constrain or mitigate. In practice, one should explore a wide variety of (sometimes incompatible) measures before deciding which ones warrant corrective action, but their narrow scope means that most new situations can only be examined after devising new measures. In this work, we present a mathematical framework that distils literature measures of bias into building blocks, hereby facilitating new combinations to cover a wide range of fairness concerns, such as classification or recommendation differences across multiple multi-value sensitive attributes (e.g., many genders and races, and their intersections). We show how this framework generalizes existing concepts and present frequently used blocks. We provide an open-source implementation of our framework as a Python library, called FairBench, that facilitates systematic and extensible exploration of potential bias concerns.

With the growing application of AI-based systems in our lives and society, there is a rising need to ensure that AI-based systems are developed and used in a responsible way. Fairness is one of the socio-technical concerns that must be addressed in AI-based systems for this purpose. Unfair AI-based systems, particularly, unfair AI-based mobile apps, can pose difficulties for a significant proportion of the global populace. This paper aims to deeply analyze fairness concerns in AI-based app reviews. We first manually constructed a ground-truth dataset including a statistical sample of fairness and non-fairness reviews. Leveraging the ground-truth dataset, we then developed and evaluated a set of machine learning and deep learning classifiers that distinguish fairness reviews from non-fairness reviews. Our experiments show that our best-performing classifier can detect fairness reviews with a precision of 94%. We then applied the best-performing classifier on approximately 9.5M reviews collected from 108 AI-based apps and identified around 92K fairness reviews. While the fairness reviews appear in 23 app categories, we found that the 'communication' and 'social' app categories have the highest percentage of fairness reviews. Next, applying the K-means clustering technique to the 92K fairness reviews, followed by manual analysis, led to the identification of six distinct types of fairness concerns (e.g., 'receiving different quality of features and services in different platforms and devices' and 'lack of transparency and fairness in dealing with user-generated content'). Finally, the manual analysis of 2,248 app owners' responses to the fairness reviews identified six root causes (e.g., 'copyright issues', 'external factors', 'development cost') that app owners report to justify fairness concerns.

Algorithmic fairness in the context of personalized recommendation presents significantly different challenges to those commonly encountered in classification tasks. Researchers studying classification have generally considered fairness to be a matter of achieving equality of outcomes between a protected and unprotected group, and built algorithmic interventions on this basis. We argue that fairness in real-world application settings in general, and especially in the context of personalized recommendation, is much more complex and multi-faceted, requiring a more general approach. We propose a model to formalize multistakeholder fairness in recommender systems as a two stage social choice problem. In particular, we express recommendation fairness as a novel combination of an allocation and an aggregation problem, which integrate both fairness concerns and personalized recommendation provisions, and derive new recommendation techniques based on this formulation. Simulations demonstrate the ability of the framework to integrate multiple fairness concerns in a dynamic way.

As recommender systems are being designed and deployed for an increasing number of socially-consequential applications, it has become important to consider what properties of fairness these systems exhibit. There has been considerable research on recommendation fairness. However, we argue that the previous literature has been based on simple, uniform and often uni-dimensional notions of fairness assumptions that do not recognize the real-world complexities of fairness-aware applications. In this paper, we explicitly represent the design decisions that enter into the trade-off between accuracy and fairness across multiply-defined and intersecting protected groups, supporting multiple fairness metrics. The framework also allows the recommender to adjust its performance based on the historical view of recommendations that have been delivered over a time horizon, dynamically rebalancing between fairness concerns. Within this framework, we formulate lottery-based mechanisms for choosing between fairness concerns, and demonstrate their performance in two recommendation domains.

Fairness Indicators is designed to support teams in evaluating and improving models for fairness concerns in partnership with the broader Tensorflow toolkit. The tool is currently actively used internally by many of our products, and is now available in BETA to try for your own use cases. Fairness Indicators enables easy computation of commonly-identified fairness metrics for binary and multiclass classifiers. At Google, it is important for us to have tools that can work on billion-user systems. Fairness Indicators will allow you to evaluate across any size of use case.