Soliciting Stakeholders’ Fairness Notions in Child Maltreatment Predictive Systems

Title of the Paper

Soliciting Stakeholders’ Fairness Notions in Child Maltreatment Predictive Systems

Paper Information

• Research Area: AI Ethics, Machine Learning Fairness, Human-Computer Interaction in Decision Systems
• Keywords: AI Fairness, Machine Learning, Algorithmic Fairness, Algorithm-Assisted Decision Making, Child Welfare

Research Background and Problem Statement

• What problems or challenges did the authors identify?

  • There are multiple technical definitions of algorithmic fairness, but these fail to reflect the diverse perspectives and needs of stakeholders in practical applications.
  • Algorithm design often neglects users' subjective perceptions of fairness, leading to fairness standards lacking legitimacy.
  • Current fairness definitions (e.g., group fairness and individual fairness) have inherent limitations and are difficult to operationalize in specific contexts.

• Why is this problem important?

  • Algorithms are widely used in high-stakes decision-making scenarios, such as criminal justice, child protection, and loan approvals, where potential unfair outcomes can significantly impact vulnerable groups.
  • Social acceptance of fairness is critical for the widespread adoption and positive societal impact of algorithmic systems.

• Research Motivation and Related Work

  • By analyzing existing literature on machine learning fairness, the authors highlight the theoretical limitations of various fairness concepts.
  • Current work on user involvement in algorithm design is insufficient to understand the complex fairness notions of different stakeholders.
  • The authors reference existing work in the child protection domain, such as the Allegheny Family Screening Tool (AFST), and emphasize the need for new methods to address fairness shortcomings in these systems.

Proposed Solution

• What methods or solutions did the authors propose?

  • The authors propose a framework for eliciting stakeholders’ subjective fairness notions, including an interactive interface and an interview protocol.
    1. Interactive Interface: Allows users to examine algorithmic predictions from both individual and group data perspectives, enabling them to articulate their fairness views.
    2. Interview Protocol: Utilizes a "think-aloud" method to explore users' feedback on fairness and algorithmic bias while interacting with the interface.

• What is innovative about this solution?

  • Integrates users’ subjective fairness notions into algorithm design, empowering users to explore and define fairness standards through the interactive interface.
  • Optimizes the process of gathering stakeholder feedback, particularly in high-stakes, sensitive domains like child protection.
  • Establishes a "macro-meso-micro" fairness framework:
    • Macro (Group Fairness): Users examine and define fairness for subgroups.
    • Meso (Case Comparison): Users compare individual cases with the dataset as a whole.
    • Micro (Individual Fairness): Users evaluate whether direct comparisons between cases are reasonable.

• What are the implementation steps and key technologies used?

  1. Design of Three Interactive Views:
     • Group Analysis View (displays algorithm performance for subgroups).
     • Case Comparison View (shows algorithm predictions and related data features for specific cases).
     • Similarity Comparison View (shows the similarity between reference cases and other cases in the dataset).
  2. Development of Interview Protocol:
     • Guide users to compare cases without algorithmic predictions first, followed by comparisons with algorithmic predictions.
     • Explore users’ fairness perspectives and record opinions on group fairness definitions (e.g., unawareness, statistical parity, equalized opportunity).
  3. Conducting User Studies:
     • Focus on child maltreatment predictive systems, selecting social workers and parents as primary stakeholders.
     • Use synthetic data to simulate real-world data while ensuring privacy protection.

Research Outcomes

• What specific findings were obtained?

  • Perspectives on Group Fairness:

    • Among respondents, "equalized opportunity" (66.7% support rate) was the most recognized fairness approach, followed by statistical parity (43.3%) and unawareness (41.7%).
    • Some respondents viewed statistical parity as a goal but believed it should not be enforced, especially for subgroups with differing base rates (e.g., children of different ages).
    • Unawareness was seen as potentially reducing or exacerbating systemic discrimination, with respondents expressing mixed opinions.

  • Feedback on Individual Fairness:

    • In case comparisons, users’ fairness standards varied even for similar cases.
    • Out of 14 case comparisons, consensus was reached in only one case (where significant differences in key attributes, such as abuse type, were present).
    • Users partially constructed narrative scenarios and partially relied on attribute importance rankings when evaluating cases.

  • Consistency and Challenges:

    • Users demonstrated high consistency when addressing fairness questions at both individual and group levels.
    • However, due to the complexity of fairness definitions and the number of data features, users faced cognitive overload when comparing highly complex cases.

• What advantages does it have compared to existing solutions?

  • Institutionalizes stakeholder feedback mechanisms into the algorithm design process.
  • Incorporates discussions of fairness and ethics into sensitive domains like child protection, rather than focusing solely on predictive accuracy.
  • Provides rich user-perspective data, including logical explanations behind fairness principles, beyond simple qualitative feedback.

• What were the experimental or evaluation results?

  • Through detailed interviews with 12 participants (social workers and parents) totaling 20.5 hours, the study collected extensive data related to algorithmic fairness.
  • The experiment demonstrated that the framework effectively captured the diversity of users’ fairness standards, highlighting how fairness definitions vary by context.

• Limitations and Future Directions

  • Limitations:
    • The sample size was limited to 12 participants and did not include all stakeholders (e.g., children, supervisory personnel).
    • The study relied on synthetic data, which may not fully reflect the complexity of real-world systems.
  • Future Directions:
    • Expand the diversity of stakeholders, especially those directly affected by algorithms.
    • Develop optimization mechanisms that incorporate the preferences of multiple stakeholders.
    • Further investigate the gap between fairness definitions and user understanding, as well as strategies to simplify trade-off decisions in algorithm design.