Towards Fairness in AI: Addressing Bias in Data Using GANs

Amirarsalan Rajabi, Ozlem O. Garibay

Abstract

Can we trust machine learning models to make fair decisions? This question becomes more relevant as these algorithms become more pervasive in many aspects of our lives and our society. While the main objective of artificial intelligence (AI) algorithms is traditionally to increase accuracy, the AI community is gradually focusing more on evaluating and developing algorithms to ensure fairness. This work explores the usefulness of adversarial learning, explicitly generative adversarial networks (GAN), in addressing the problem of fairness. We show that the proposed model is able to produce synthetic tabular data to augment the original dataset in order to improve demographic parity, while maintaining data utility. In doing so, our work increases algorithmic fairness while maintaining accuracy.
BibTeX
@inproceedings{rajabi2021towards,
  author = {Amirarsalan Rajabi and Ozlem O. Garibay},
  title = {{Towards Fairness in AI: Addressing Bias in Data Using GANs}},
  booktitle = {HCI},
  year = {2021},
  doi = {10.1007/978-3-030-90963-5\_39},
}

Loading related content...

Figure