Fairness GAN: Generating Datasets with Fairness Properties using a Generative Adversarial Network

Samuel C. Hoffman
Samuel C. Hoffman
Kush Raj Varshney
Kush Raj Varshney

IBM Journal of Research and Development, pp. 3:1-3:9, 2019.

Cited by: 9|Views12
EI WOS
Weibo:
We propose a new debiasing approach based on generative adversarial networks

Abstract:

We introduce the Fairness GAN (generative adversarial network), an approach for generating a dataset that is plausibly similar to a given multimedia dataset, but is more fair with respect to protected attributes in decision making. We propose a novel auxiliary classifier GAN that strives for demographic parity or equality of opportunity a...More

Code:

Data:

0
Introduction
  • Automated essay scoring in high-stakes educational assessment [1, 2] and automated employment screening based on voice and video [3, 4] are examples of decision making with multimedia inputs supported by machine learning algorithms that raise concerns about perpetuating and scaling unwanted bias present in historical data and running afoul of laws, e.g., Title VI and Title VII of the Civil Rights Act of 1964 in the United States.
  • The objective of the Fairness GAN is to take a given real dataset ðCreal; Xreal; YrealÞ and learn to generate debiased data ðXfake; YfakeÞ such that the joint distribution of the features and outcome of the generated data is close to that of the real data while yielding decisions that have either demographic parity or equality of opportunity.
  • LRSX 1⁄4 E1⁄2logP ðSX 1⁄4 real j Xrealފ LFSX 1⁄4 E1⁄2logP ðSX 1⁄4 fake j Xfakeފ: The authors include a pair of class-conditioned losses to add structure to the GAN and help with training and generating plausible images.
Highlights
  • Automated essay scoring in high-stakes educational assessment [1, 2] and automated employment screening based on voice and video [3, 4] are examples of decision making with multimedia inputs supported by machine learning algorithms that raise concerns about perpetuating and scaling unwanted bias present in historical data and running afoul of laws, e.g., Title VI and Title VII of the Civil Rights Act of 1964 in the United States
  • We propose a new debiasing approach based on generative adversarial networks (GANs) [8]
  • We focus on the category “power outlet” as it is known to have differential recognition performance on submitters from Great Britain, and Canada and the United States [28]
  • Because of the small number of real samples in the soccer dataset, it is impractical to train a classifier with a similar architecture to the GAN discriminator like we do with the other datasets
  • In this article, we have examined fairness in the scenario of binary classification with multimedia features and developed a GAN-based preprocessing approach to improve demographic parity or equality of opportunity by learning to generate a fairer dataset in the original input feature space
Results
  • The last term (LRDP ) is to train the discriminator to predict the conditioning variable from the outcome Y alone for real data.
  • The last term is aimed at reducing the ability of the discriminator to correctly predict the conditioning variable from the outcome Y alone for generated dataset.
  • While the work is geared towards high-dimensional image data and apart from demographic parity the architecture supports equality of opportunity.
  • Because of the small number of real samples in the soccer dataset, it is impractical to train a classifier with a similar architecture to the GAN discriminator like the authors do with the other datasets.
  • The authors perform a random 90/10 training/testing partition of the data (70/30 partition for soccer) and use the training set for independently learning two GANs: one with demographic parity loss and one with equality of opportunity loss.
  • Empirical results Table 1 presents the different error rates for the three classifiers conditioned on the protected attribute, the overall unconditional error rate, and the values of demographic parity loss and equality of opportunity loss for all four datasets.
  • Figure 1 presents the receiver operating characteristics (ROCs) for the case without debiasing and the demographic parity Fairness GAN, broken down by the two values of the protected attribute.
  • Conclusion In this article, the authors have examined fairness in the scenario of binary classification with multimedia features and developed a GAN-based preprocessing approach to improve demographic parity or equality of opportunity by learning to generate a fairer dataset in the original input feature space.
Conclusion
  • The authors use the proposed algorithm, the first application of GANs to algorithmic fairness, to process several attributed image datasets with varied properties, outcome variables, Eigenfaces from the CelebA dataset.
  • Fairness definition [12], pursuing the Gumbel-Softmax trick for discrete outcome variables [31, 32], and formulating a Fairness GAN for continuous protected attributes [35].
  • The work so far illuminates several directions for future research, e.g., considering other modalities of multimedia data in addition to images, adding the equality of odds
Summary
  • Automated essay scoring in high-stakes educational assessment [1, 2] and automated employment screening based on voice and video [3, 4] are examples of decision making with multimedia inputs supported by machine learning algorithms that raise concerns about perpetuating and scaling unwanted bias present in historical data and running afoul of laws, e.g., Title VI and Title VII of the Civil Rights Act of 1964 in the United States.
  • The objective of the Fairness GAN is to take a given real dataset ðCreal; Xreal; YrealÞ and learn to generate debiased data ðXfake; YfakeÞ such that the joint distribution of the features and outcome of the generated data is close to that of the real data while yielding decisions that have either demographic parity or equality of opportunity.
  • LRSX 1⁄4 E1⁄2logP ðSX 1⁄4 real j Xrealފ LFSX 1⁄4 E1⁄2logP ðSX 1⁄4 fake j Xfakeފ: The authors include a pair of class-conditioned losses to add structure to the GAN and help with training and generating plausible images.
  • The last term (LRDP ) is to train the discriminator to predict the conditioning variable from the outcome Y alone for real data.
  • The last term is aimed at reducing the ability of the discriminator to correctly predict the conditioning variable from the outcome Y alone for generated dataset.
  • While the work is geared towards high-dimensional image data and apart from demographic parity the architecture supports equality of opportunity.
  • Because of the small number of real samples in the soccer dataset, it is impractical to train a classifier with a similar architecture to the GAN discriminator like the authors do with the other datasets.
  • The authors perform a random 90/10 training/testing partition of the data (70/30 partition for soccer) and use the training set for independently learning two GANs: one with demographic parity loss and one with equality of opportunity loss.
  • Empirical results Table 1 presents the different error rates for the three classifiers conditioned on the protected attribute, the overall unconditional error rate, and the values of demographic parity loss and equality of opportunity loss for all four datasets.
  • Figure 1 presents the receiver operating characteristics (ROCs) for the case without debiasing and the demographic parity Fairness GAN, broken down by the two values of the protected attribute.
  • Conclusion In this article, the authors have examined fairness in the scenario of binary classification with multimedia features and developed a GAN-based preprocessing approach to improve demographic parity or equality of opportunity by learning to generate a fairer dataset in the original input feature space.
  • The authors use the proposed algorithm, the first application of GANs to algorithmic fairness, to process several attributed image datasets with varied properties, outcome variables, Eigenfaces from the CelebA dataset.
  • Fairness definition [12], pursuing the Gumbel-Softmax trick for discrete outcome variables [31, 32], and formulating a Fairness GAN for continuous protected attributes [35].
  • The work so far illuminates several directions for future research, e.g., considering other modalities of multimedia data in addition to images, adding the equality of odds
Tables
  • Table1: Different error rates for the three classifiers conditioned on the protected attribute, the overall unconditional error rate, and the values of demographic parity loss and equality of opportunity loss for all four datasets
Download tables as Excel
Reference
  • M. D. Shermis, “State-of-the-art automated essay scoring: Competition, results, and future directions from a United States demonstration,” Assessing Writing, vol. 20, pp. 53–76, Apr. 2014.
    Google ScholarLocate open access versionFindings
  • L. Perelman, “When ‘the state of the art’ is counting words,” Assessing Writing, vol. 21, pp. 104–11, Jul. 2014.
    Google ScholarLocate open access versionFindings
  • A. Shahani, “Now algorithms are deciding whom to hire, based on voice,” 2015. [Online]. Available: htt_ps://www_.npr.org/sections/ alltechconsidered/2015/03/ 23/394827451/now-algorithms-aredeciding-whom-to-hire-based-on-voice
    Google ScholarFindings
  • S. Chandler, “The AI chatbot will hire you now,” 2017. [Online]. Available: htt_ps://www_.wired.com/story/the-ai-chatbot-will-hireyou-now/.
    Google ScholarFindings
  • B. A. Williams, C. F. Brooks, and Y. Shmargad, “How algorithms discriminate based on data they lack: Challenges, solutions, and policy implications,” J. Inf. Policy, vol. 8, pp. 78–115, 2018.
    Google ScholarLocate open access versionFindings
  • B. d’Alessandro, C. O’Neil, and T. LaGatta, “Conscientious classification: A data scientist’s guide to discrimination-aware classification,” Big Data, vol. 5, no. 2, pp. 120–34, 2017.
    Google ScholarLocate open access versionFindings
  • F. Kamiran and T. Calders, “Data preprocessing techniques for classification without discrimination,” Knowl. Inf. Syst., vol. 33, no. 1, pp. 1–33, 2012.
    Google ScholarLocate open access versionFindings
  • I. Goodfellow, “NIPS 2016 tutorial: Generative adversarial networks,” 2017, arXiv:1701.00160.
    Findings
  • A. Odena, C. Olah, and J. Shlens, “Conditional image synthesis with auxiliary classifier GANs,” in Proc. Int. Conf. Mach. Learn., 2017, pp. 2642–51.
    Google ScholarLocate open access versionFindings
  • H. Edwards and A. Storkey, “Censoring representations with an adversary,” in Proc. Int. Conf. Learn. Rep., 2016.
    Google ScholarLocate open access versionFindings
  • A. Beutel, J. Chen, Z. Zhao, et al., “Data decisions and theoretical implications when adversarially learning fair representations,” in Proc. Workshop Fairness, Accountability, Transparency Mach. Learn., 2017.
    Google ScholarLocate open access versionFindings
  • B. H. Zhang, B. Lemoine, and M. Mitchell, “Mitigating unwanted biases with adversarial learning,” in Proc. AAAI/ACM Conf. Artif. Intell., Ethics, Soc., 2018, pp. 335–340.
    Google ScholarLocate open access versionFindings
  • D. Madras, E. Creager, T. Pitassi, et al., “Learning adversarially fair and transferable representations,” in Proc. Int. Conf. Mach. Learn., July 2018, pp. 3384–3393.
    Google ScholarLocate open access versionFindings
  • C. Wadsworth, F. Vera, and C. Piech, “Achieving fairness through adversarial learning: An application to recidivism prediction,” in Proc. Workshop Fairness, Accountability, Transparency Mach. Learn., 2018.
    Google ScholarLocate open access versionFindings
  • T. Adel, I. Valera, Z. Ghahramani, et al., “One-network adversarial fairness,” in Proc. 33rd AAAI Conf. Artif. Intell., 2019, pp. 2412–2420.
    Google ScholarLocate open access versionFindings
  • Y. Elazar and Y. Goldberg, “Adversarial removal of demographic attributes from text data,” in Proc. Conf. Empirical Methods Natural Lang. Process., 2018, pp. 11–21.
    Google ScholarLocate open access versionFindings
  • G. Perarnau, J. Van de Weijer, B. Raducanu, et al., “Invertible conditional GANs for image editing,” in Proc. NIPS Workshop Adversarial Training, 2016.
    Google ScholarLocate open access versionFindings
  • J.-Y. Zhu, T. Park, P. Isola, et al., “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Proc. IEEE Int. Conf. Comput. Vis., 2017, pp. 2242–51.
    Google ScholarLocate open access versionFindings
  • Y. Lu, Y.-W. Tai, and C.-K. Tang, “Conditional CycleGAN for attribute guided face image generation,” 2017, arXiv:1705.09966.
    Findings
  • Z. Liu, P. Luo, X. Wang, et al., “Deep learning face attributes in the wild,” in Proc. IEEE Int. Conf. Comput. Vis., 2015, pp. 3730–8.
    Google ScholarLocate open access versionFindings
  • R. Silberzahn, E. L. Uhlmann, D. P. Martin, et al., “Many analysts, one dataset: Making transparent how variations in analytical choices affect results,” Adv. Methods Pract. Psychol. Sci., vol. 1, no. 3, pp. 337–356, 2017.
    Google ScholarLocate open access versionFindings
  • V. Dumoulin, J. Shlens, and M. Kudlur, “A learned representation for artistic style,” in Proc. Int. Conf. Learn. Rep., 2017.
    Google ScholarLocate open access versionFindings
  • T. Miyato and M. Koyama, “cGANs with projection discriminator,” in Proc. Int. Conf. Learn. Rep., 2018.
    Google ScholarLocate open access versionFindings
  • T. Miyato, T. Kataoka, M. Koyama, et al., “Spectral normalization for generative adversarial networks,” in Proc. Int. Conf. Learn. Rep., 2018.
    Google ScholarLocate open access versionFindings
  • J. H. Lim and J. C. Ye, “Geometric GAN,” 2017, arXiv:1705.02894.
    Findings
  • D. Xu, S. Yuan, L. Zhang, et al., “FairGAN: Fairness-aware generative adversarial networks,” in Proc. IEEE Int. Conf. Big Data, Dec. 2018, pp. 570–575.
    Google ScholarLocate open access versionFindings
  • M. B€ohlen, V. Chandola, and A. Salunkhe, “Server, server in the cloud. Who is the fairest in the crowd?,” 2017, arXiv:1711.08801.
    Findings
  • TailSpectrum, “Quickdraw with Google didn’t recognise my drawing of a power outlet because I drew a 3-pinned rectangular plug,” 2016. [Online]. Available: htt_ps://www_.reddit.com/r/ britishproblems/comments/5deygp/ quickdraw_with_google_didnt_recognise_my_drawing
    Google ScholarFindings
  • I. Gulrajani, F. Ahmed, M. Arjovsky, et al., “Improved training of wasserstein GANs,” in Proc. Adv. Neural Inf. Process. Syst., 2017, pp. 5769–79.
    Google ScholarLocate open access versionFindings
  • K. He, X. Zhang, S. Ren, et al., “Deep residual learning for image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 770–78.
    Google ScholarLocate open access versionFindings
  • E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,” in Proc. Int. Conf. Learn. Rep., 2017.
    Google ScholarLocate open access versionFindings
  • C. J. Maddison, A. Mnih, and Y. W. Teh, “The concrete distribution: A continuous relaxation of discrete random variables,” in Pro. Int. Conf. Learn. Rep., 2017.
    Google ScholarLocate open access versionFindings
  • M. Turk and A. Pentland, “Eigenfaces for recognition,” J. Cogn. Neurosci., vol. 3, no. 1, pp. 71–86, 1991.
    Google ScholarLocate open access versionFindings
  • S. A. Friedler, C. Scheidegger, S. Venkatasubramanian, et al., “A comparative study of fairness-enhancing interventions in machine learning,” in Proc. ACM Conf. Fairness, Accountability Transparency, Jan. 2019, pp. 329–338.
    Google ScholarLocate open access versionFindings
  • G. Louppe, M. Kagan, and K. Cranmer, “Learning to pivot with adversarial networks,” in Proc. Adv. Neural Inf. Process. Syst., 2017, pp. 981–90. September 24, 2019
    Google ScholarLocate open access versionFindings
  • Prasanna Sattigeri IBM Research, Yorktown Heights, NY 10598 USA (psattig@us.ibm.com). Dr. Sattigeri received a Ph.D. degree in electrical engineering from the Arizona State University, Tempe, AZ, USA, in 2014. He is currently a Research Staff Member with IBM Research, Yorktown Heights, NY, USA. His current work focuses on exploring structure in the data using deep generative models and developing algorithms that are data-efficient. His broad research interests include machine learning and signal processing. His research also involves developing theory and practical systems for machine learning applications that demand constraints such as robustness, fairness, and interpretability.
    Google ScholarLocate open access versionFindings
  • Samuel C. Hoffman IBM Research, Yorktown Heights, NY 10598 USA (shoffman@ibm.com). Mr. Hoffman received a B.S. degree in computer science and mechanical engineering from Cornell University, Ithaca, NY, USA, in 2017. After graduation, he has been working as a Research Software Engineer with IBM Research, Yorktown Heights, NY, USA. His research interests include deep learning, generative modeling, and algorithmic fairness.
    Google ScholarLocate open access versionFindings
  • Kush R. Varshney IBM Research, Yorktown Heights, NY 10598 USA (krvarshn@us.ibm.com). Dr. Varshney received a B.S. degree from Cornell University, Ithaca, NY, USA, in 2004, and S.M. and Ph.D. degrees from the Massachusetts Institute of Technology, Cambridge, MA, USA, in 2006 and 2010, respectively. He is currently a Principal Research Staff Member and a Manager with IBM Research, Thomas J. Watson Research Center, Yorktown Heights, NY, USA, where he leads the Machine Learning group in the Foundations of Trusted AI Department. He is the founding CoDirector of the IBM Science for Social Good initiative. He is a Senior Member of the IEEE and a member of the Partnership on AI’s SafetyCritical AI expert group.
    Google ScholarLocate open access versionFindings
Your rating :
0

 

Tags
Comments