[Paper Review] Discriminative but Not Discriminatory: A Comparison of Fairness Definitions under Different Worldviews.

Samuel Yeom, Michael Carl Tschantz|arXiv (Cornell University)|Aug 26, 2018

Gender Diversity and Inequality10 citations

TL;DR

This paper compares three fairness definitions—demographic parity, equalized odds, and predictive parity—within distinct worldviews that assume varying degrees of data bias. It argues that demographic parity and equalized odds are justified under specific worldviews to prevent disparity amplification, while predictive parity fails to prevent large inter-group disparities, leading to a proposed new fairness notion aligned with a more realistic worldview.

ABSTRACT

We mathematically compare three competing definitions of group-level nondiscrimination: demographic parity, equalized odds, and predictive parity. Using the theoretical framework of Friedler et al., we study the properties of each definition under various worldviews, which are assumptions about how, if at all, the observed data is biased. We argue that different worldviews call for different definitions of fairness, and we specify the worldviews that, when combined with the desire to avoid a criterion for discrimination that we call disparity amplification, motivate demographic parity and equalized odds. In addition, we show that predictive parity is insufficient for avoiding disparity amplification because it allows an arbitrarily large inter-group disparity. Finally, we define a worldview that is more realistic than the previously considered ones, and we introduce a new notion of fairness that corresponds to this worldview.

Motivation & Objective

• To analyze how different assumptions about data bias (worldviews) influence the appropriateness of fairness definitions in machine learning.
• To identify which fairness definitions—demographic parity, equalized odds, or predictive parity—are most suitable under specific worldviews.
• To demonstrate that predictive parity is insufficient to prevent disparity amplification due to its tolerance for large inter-group disparities.
• To introduce a new fairness notion grounded in a more realistic worldview that better reflects real-world data biases.

Proposed method

• Uses the theoretical framework of Friedler et al. to model fairness definitions under distinct worldviews, each representing assumptions about data bias.
• Defines 'disparity amplification' as a criterion to evaluate whether a fairness definition prevents worsening of existing group disparities.
• Analyzes the mathematical properties of demographic parity, equalized odds, and predictive parity under each worldview to assess their discriminatory implications.
• Introduces a new worldview that better reflects real-world data biases, such as historical inequities and measurement errors.
• Derives a new fairness criterion that aligns with this realistic worldview, ensuring robustness against disparity amplification.

Experimental results

Research questions

• RQ1 Under what worldviews are demographic parity and equalized odds justified as fairness criteria?
• RQ2 Why does predictive parity fail to prevent disparity amplification despite being considered fair in some contexts?
• RQ3 What new fairness definition emerges when adopting a more realistic worldview that accounts for real-world data biases?
• RQ4 How do different assumptions about data bias affect the validity of standard fairness definitions?

Key findings

• Demographic parity and equalized odds are justified under specific worldviews that assume certain types of data bias, particularly when avoiding disparity amplification.
• Predictive parity allows arbitrarily large inter-group disparities, making it insufficient to prevent disparity amplification.
• A new fairness notion is proposed that aligns with a more realistic worldview, offering stronger protection against disparity amplification.
• The study shows that fairness definitions must be selected based on underlying assumptions about data bias, not applied universally.