Perceptions of the Fairness Impacts of Multiplicity in Machine Learning

Research Background and Problem

• Identified Issues or Challenges: The authors explore the impact of "diversity" on the fairness of machine learning (ML), where diversity refers to the existence of multiple different models with comparable predictive accuracy. This diversity can lead to the same input data being predicted differently by different models, potentially creating risks to procedural fairness. However, current literature has not investigated whether "stakeholders" also perceive this risk.
• Significance: When ML is applied to high-stakes decisions (e.g., social services, criminal justice, and healthcare), such arbitrariness can have substantial impacts on individuals, such as losing job opportunities or access to healthcare. This directly affects public trust and acceptance of ML systems.
• Research Motivation and Related Work: Existing studies from philosophical and computer science perspectives suggest that diversity may pose risks to fairness. However, there is a lack of direct dialogue with ordinary stakeholders (e.g., decision subjects). Given that the outputs of ML systems affect public interests, understanding their perceptions is crucial.

Solution

• Method or Solution: Through a survey study, the authors measured the perceptions of the general public (i.e., decision recipients) regarding the impact of diversity on fairness in ML systems and their preferences for addressing diversity issues.
• Innovative Contribution: This is the first systematic investigation into non-technical stakeholders' views on the fairness implications of diversity, bridging the gap between theoretical assumptions and public perceptions.
• Implementation Steps:
  1. Conduct a preliminary study to select several tasks.
  2. Use a main study to investigate how diversity affects participants' perceptions of ML fairness.
  3. Collect participants' preferences and compare six diversity resolution methods (e.g., randomization, human expert intervention).
  4. Analyze response patterns using qualitative and quantitative methods, and examine how task characteristics (e.g., risk and reward frameworks) moderate participants' responses.

Research Findings

• Specific Findings:
  1. Ordinary stakeholders generally do not perceive diversity as a significant threat to the fairness of ML systems, but they exhibit clear preferences for resolving diversity issues.
  2. The most preferred approach is human expert intervention, while the least preferred is randomization.
  3. Preferences vary depending on task characteristics; for instance, in high-risk tasks, participants favor human decision-making.
• Comparison with Existing Solutions:
  • Current ML practices often directly select a single well-performing model while ignoring diversity, but this approach (ignore) is perceived as unfair by the public.
  • Philosophers' proposal of randomization as a fairness solution is also not accepted by the public.
• Experimental or Evaluation Results:
  • Preference for "human decision-making" (marginal mean of 0.795, significantly higher than the random expectation value of 0.5).
  • In high-risk tasks, there is a stronger preference for "human" or "complex models" (e.g., ensemble methods), while randomization is more tolerable in low-risk tasks.
• Limitations and Future Directions:
  1. Participants' understanding of diversity could be further improved through education, as some results may be constrained by their technical knowledge.
  2. The current study is based on hypothetical tasks; future research could focus on real-world scenarios to collect data with higher ecological validity.
  3. The reasons behind participants' preferences, particularly their strong aversion to randomization (which contradicts philosophical recommendations), warrant further exploration.

Conclusion

This study challenges the assumptions of current ML practices and philosophical recommendations, particularly in addressing the issue of predictive diversity. While the presence of diversity does not significantly reduce participants' perceptions of fairness, they exhibit clear preferences for resolving diversity issues. This suggests that ML developers need to address diversity issues more transparently and purposefully, while considering stakeholder involvement and perceptions in system design. The research provides a new perspective on the societal dimensions of ML fairness and lays a foundation for enhancing public trust and acceptance of algorithmic decision-making in the future.