Learning Disentangled Representations for Fairness with Limited Demographics

Shenyu Lu, Xiaoqian Wang

Keywords: Fair representation learning, Disentanglement

TL;DR: We propose a method that learns a fair representation with limited sensitive information available.

Abstract: Fair representation learning is a promising way to mitigate discrimination in downstream tasks. Many existing fair representation learning methods require access to sensitive information, but the collection of sensitive information is often difficult and even involves privacy issues. Additionally, a model trained to be fair with respect to one sensitive attribute may not ensure fairness for other sensitive groups. Thus, how to flexibly address fairness issues when we have limited access to sensitive information is a challenging problem. In this work, we answer this question: ``given limited sensitive information, can we learn a representation to be fair w.r.t. varying sensitive groups?'' To achieve this, we propose a novel two-step framework. We first learn a disentangled representation by employing Non-linear Independent Component Analysis (Nonlinear ICA). Second, we remove sensitive information in the latent space to obtain fair representation. The learned representation can be easily adapted to be fair w.r.t different sensitive groups and to be used for different downstream tasks without re-training. Among the entire process, only a small portion of sensitive information is required in the second step to learn a fair representation. We compare with methods that require different amounts of sensitive information on real-world images and tabular datasets. We empirically demonstrate the utility and flexibility of our approach, and our method is capable of achieving improved fairness results in various tasks.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8361