Fairness-enhancing mixed effects deep learning improves fairness on in- and out-of-distribution clustered (non-iid) data
Primary Area: societal considerations including fairness, safety, privacy
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.
Keywords: Fairness-enhancing models, adversarial debiasing, mixed effects deep learning, out of distribution generalization
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: A novel framework integrates mixed effects deep learning and adversarial debiasing, boosting performance, generalization, while promoting equality-of-odds fairness across all tested sensitive variables like race, age, sex, and marital-status.
Abstract: Traditional deep learning (DL) suffers from two core problems. First, it assumes training samples are independent and identically distributed (iid), however, in many real-world datasets samples are grouped by measurements made on the same sample (e.g., study participant, cell, tissue) violating this assumption. On such clustered data, traditional DL suffers from reduced prediction performance, lack of generalization, poor interpretability, and cluster confounding causing Type 1 and 2 errors. Second, traditional model fitting prioritizes only overall training data accuracy, which is biased towards the most common group, with often unfair, lower accuracy on samples from underrepresented groups. When DL is used to guide critical decisions (e.g., loan approvals or determining health insurance rates) such biases can significantly impact one’s quality of life. To address both of these challenges simultaneously, we introduce a fairness-enhancing mixed effects deep learning (MEDL) framework. MEDL separately quantifies cluster-invariant fixed effects (FE) and cluster-specific random effects (RE) through the introduction of: 1) a cluster adversary which encourages the learning of cluster-invariant FE, 2) a Bayesian neural network which quantifies the RE, and a mixing function combining the FE an RE into a mixed-effect prediction. We marry this MEDL with adversarial debiasing, which promotes equality-of-odds fairness across FE, RE, and ME predictions for fairness-sensitive variables. An empirical evaluation spanning three datasets: two from the census/finance sector and one from the healthcare sector is performed. The former focuses on income classification, while the latter is a healthcare dataset that predicts forthcoming hospitalization duration, a regression task. The proposed framework boosts fairness across all sensitive variables—increasing fairness up to 82% for age, 43% for race, 86% for sex, and 27% for marital-status. While enhancing fairness, our method also preserves the intrinsic improved performance and interpretability advantages of MEDL. Moreover, the proposed method is agnostic to dataset type and task (regression or classification) with broad potential applicability by the community. To facilitate these benefits and further community extension, we have made our implementation available through GitHub.
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: 8710
Loading