Amirarsalan Rajabi

Amirarsalan Rajabi

AI, Machine Learning

Fair Bilevel Neural Network (FairBiNN): On Balancing Fairness and Accuracy via Stackelberg Equilibrium

Mehdi Yazdani-Jahromi, Ali Khodabandeh Yalabadi, Amirarsalan Rajabi, Aida Tayebi, Ivan Garibay, Ozlem Ozmen Garibay

Advances in Neural Information Processing Systems (NeurIPS 2024), 2024

Paper arXiv Code BibTeX

TL;DR Fairness and accuracy are treated as two players in a Stackelberg game, each owning its own layers of one network. The bilevel solution is provably no worse than the usual Lagrangian (regularization) approach, and in practice it reaches a better accuracy–fairness trade-off.

SchematicFairBiNN splits one network between two players. The accuracy player (leader) owns most layers and minimizes BCE. The fairness player (follower) owns a small block of fairness layers and minimizes the demographic-parity loss. Training alternates between their updates, which is a Stackelberg game.

Abstract

The persistent challenge of bias in machine learning models necessitates robust solutions to ensure parity and equal treatment across diverse groups, particularly in classification tasks. Current methods for mitigating bias often result in information loss and an inadequate balance between accuracy and fairness. To address this, we propose a novel methodology grounded in bilevel optimization principles. Our deep learning-based approach concurrently optimizes for both accuracy and fairness objectives, achieving Pareto optimal solutions while mitigating bias in the trained model. Theoretical analysis shows that the upper bound on the loss incurred by this method is less than or equal to the loss of the Lagrangian approach. We demonstrate the efficacy of our model on tabular datasets such as UCI Adult and Heritage Health, outperforming state-of-the-art fairness methods.

Method

The FairBiNN architecture. The accuracy player's layers (θa) are trained on binary cross-entropy. A block of fairness layers (θf) inside the network is trained separately on a demographic-parity loss, with its own optimizer.
The FairBiNN architecture. The accuracy player’s layers (θa) are trained on binary cross-entropy. A block of fairness layers (θf) inside the network is trained separately on a demographic-parity loss, with its own optimizer.

The accuracy player is the leader: it solves the upper-level problem, taking into account how the fairness player will respond. The fairness player is the follower: it minimizes the fairness loss given the leader’s parameters. Training alternates between the two updates. Under the paper’s assumptions, the resulting Stackelberg equilibrium is Pareto optimal, and its loss is bounded by that of the Lagrangian formulation.

Results

(a) UCI Adult
(a) UCI Adult
(b) Heritage Health
(b) Heritage Health

Accuracy compared with statistical demographic parity for FairBiNN (red diamonds) and benchmark methods. The best region is the bottom right: high accuracy and low parity difference. FairBiNN gives the best trade-off on both datasets.

(a) Without the fairness layers, z
(a) Without the fairness layers, z
(b) With the fairness layers, z̃
(b) With the fairness layers, z̃

CelebA (target: Attractive): t-SNE of the representation, colored by gender. With the fairness layers, the representation no longer clusters by gender.

Citation

@inproceedings{yazdanijahromi2024fairbinn,
  title     = {Fair Bilevel Neural Network (FairBiNN): On Balancing Fairness and Accuracy via Stackelberg Equilibrium},
  author    = {Yazdani-Jahromi, Mehdi and Khodabandeh Yalabadi, Ali and Rajabi, Amirarsalan and Tayebi, Aida and Garibay, Ivan and Ozmen Garibay, Ozlem},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {37},
  pages     = {105780--105818},
  year      = {2024}
}