Effect of Information Presentation on Fairness Perceptions of Machine Learning Predictors

Title of the Paper

Effect of Information Presentation on Fairness Perceptions of Machine Learning Predictors

Paper Information

• Domain: AI Fairness and Human-Computer Interaction
• Keywords: Artificial Intelligence, Fairness, Transparency, Crowdsourcing, Machine Learning, Predictor Selection, General Users, Data Visualization, AI, ML.

Research Background and Problem

• What problems or challenges did the authors identify?

  • While Artificial Intelligence (AI) technologies have improved the efficiency of many decision-making processes, they have also raised significant concerns regarding fairness and transparency. For example, gender bias in AI recruitment tools or racial bias in criminal justice prediction systems.
  • The general public is considered an important stakeholder in the design and evaluation of algorithms, but many existing Machine Learning (ML) explanation tools are designed primarily for experts, without adequately addressing the comprehension needs of non-experts.
  • There is insufficient understanding in the literature regarding how information presentation methods (e.g., visualization techniques) influence public perceptions of fairness in AI systems.

• Why is this problem important?

  • For society at large, the fairness of AI directly impacts trust and acceptance.
  • Non-expert users may have perspectives that differ from technical experts; excluding their participation in algorithm design can lead to non-representative systems, exacerbating potential unfairness.

• Motivation and Related Work

  • The authors reviewed prior studies in the fields of fairness and public perception (e.g., different definitions of individual fairness and group fairness), highlighting that perceived fairness is influenced by multiple factors such as information presentation methods, education level, and gender differences.
  • Previous literature in HCI and AI has discussed the impact of information presentation but has not specifically tested the role of visualization techniques and contextual data in fairness evaluations.

Solution

• Research Methodology:

  • An online crowdsourcing study was designed to explore the impact of different information presentation methods (text vs. scatterplot) and outcome variable display (shown/not shown) on public perceptions of fairness.
  • Core variables included predictor types (e.g., demographic indicators vs. validation variables), scenarios (criminal recidivism risk vs. loan approval), presentation methods, and demographic characteristics.

• Key Innovations:

  1. Integration of text and scatterplot visualization methods to jointly study their impact on fairness evaluations.
  2. Exploration of contextual and demographic factors (e.g., gender and education level) in assessing the fairness of predictors.
  3. Design of nested experiments to validate the applicability of different algorithmic predictors across two scenarios.

• Implementation Steps:

  1. Using two real-world datasets (COMPAS recidivism dataset and loan dataset), predictors were defined, including demographic, domain-specific, and validation-related variables.
  2. Participants evaluated whether each predictor should be included in the algorithm and rated its fairness using different visualization and variable presentation methods.
  3. Background data were collected from participants, including gender, education level, and graph interpretation skills, and analyzed for relationships with perceived fairness.

Research Findings

• Specific Findings:

  • Text-based presentation methods were more likely to lead participants to perceive predictors as fair compared to scatterplots.
  • When outcome variables were displayed, fairness ratings in the criminal recidivism scenario significantly decreased, whereas no significant changes were observed in the loan scenario.
  • Validation predictors (e.g., "height and dominant hand") were generally perceived by the public as unfair.

• Advantages Compared to Existing Solutions:

  • The study incorporated the interactive effects of context and data presentation methods into the research framework, laying the groundwork for designing more effective user-friendly AI evaluation tools.
  • It highlighted the importance of demographic characteristics (e.g., gender, education level) in public perceptions of fairness, addressing gaps in existing literature.

• Experimental or Evaluation Results:

  • Participants generally rated the fairness of predictors in the criminal recidivism scenario higher than those in the loan scenario.
  • The complexity of data visualization directly impacted audience confidence; the "scatterplot-outcome" combination condition resulted in the lowest confidence among participants.
  • Highly educated individuals had higher fairness expectations and were more likely to perceive information as insufficient; women tended to perceive algorithmic predictors as less fair.

• Limitations and Future Directions:

  • The study was limited to text and scatterplot presentation methods; future research could explore other visualization formats.
  • The datasets used were restricted to the U.S., and further evaluation of cultural/geographical influences on fairness perceptions is needed.
  • Future research could investigate the practical effects of user participation across multiple stages of algorithm development (e.g., predictor review, model deployment).

Conclusion and Recommendations

• There is a need to develop explanatory tools focused on general public users, avoiding overly complex presentation formats that may distort research outcomes.
• Future studies should conduct cross-validation across diverse scenarios and visualization techniques to establish a foundation for building inclusive and diverse AI systems.