Private but Biased? Exploring Fairness in Federated Recommendation

TMLR Paper9523 Authors

Abstract:

Bias and fairness are central concerns in machine learning, particularly in recommendation systems that may reinforce gender, age, or occupation stereotypes.
In parallel, federated recommendation has emerged as a privacy-preserving alternative to centralized systems, particularly in cross-device settings where data remains on user devices.
Despite extensive studies of bias in centralized recommendation, its behavior under federated training remains largely underexplored.
Indeed, the transition to a decentralized architecture introduces additional sources of statistical skew and uneven representation across users, making the impact of federated learning on bias dynamics unclear.
Furthermore, most existing bias mitigation techniques rely on sharing sensitive user or item attributes, which conflicts with the privacy constraints inherent to federated learning.
In this work, we investigate how federated training influences the emergence of gender and user activity bias in cross-device federated recommendation systems.
We further adapt an existing bias mitigation approach to the federated setting and propose a privacy-aware framework for bias mitigation that does not require sharing sensitive attributes.
Our results show that, under certain conditions, federated training can introduce less bias than its centralized counterpart.
Across datasets and model families, we observe a consistent reduction in gender-based bias under FL settings, while popularity (activity) bias exhibits model-dependent behavior and can increase in graph-based user-expansion methods.
Moreover, we demonstrate that effective bias mitigation is feasible in federated recommendation while preserving user privacy. Our Source code, datasets and trained models are available at \url{https://anonymous.4open.science/r/Bias-and-cross-device-Federated-recommendation-50D5/}

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Audra_McMillan1

Submission Number: 9523