Towards a Non-Ideal Methodological Framework for Responsible ML

Document Title

Towards a Non-Ideal Methodological Framework for Responsible ML

Document Information

• Topic Area: Methodological exploration of Responsible Machine Learning (RML)
• Keywords: Responsible Machine Learning, fairness, machine learning, justice, ideal theory, non-ideal theory, ML practitioners
• Conference and Publication Information: CHI Conference 2024 (CHI '24)

Research Background and Issues

• Problems and Challenges:

  1. Current research on Responsible Machine Learning (RML) predominantly focuses on technical aspects, but it remains unclear how technical practitioners can incorporate these principles into concrete operations.
  2. ML practitioners (such as developers, engineers, data scientists) lack systematic documentation and research on their assumptions, choices, and methods when applying RML, leaving many implicit practices underexplored.
  3. Existing RML approaches are largely based on "ideal theory," which proposes abstract fairness principles through simplified assumptions, overlooking the complexities of "non-ideal conditions" in real-world environments.

• Research Significance: Understanding the methodological characteristics of how technical practitioners approach RML can provide new insights for improving RML theoretical frameworks and practices, while fostering multi-stakeholder collaboration.

• Research Motivation: To shift RML methods from idealized approaches to being sensitive to complex real-world contexts and explore how practitioners establish methodological pathways in practical cases.

Solution

• Proposed Method: The authors propose a new methodological framework based on "non-ideal theory," which:

  1. Emphasizes diagnosing and systematically documenting imperfect conditions in non-ideal contexts.
  2. Introduces a mapping mechanism to connect "abstract values" (e.g., fairness) with pathways to achieve goals and specific technical actions.
  3. Integrates and visualizes methodological choices, assumptions, and their potential impacts to support collaboration between technical and non-technical stakeholders.

• Innovative Aspects:

  1. Introduces "non-ideal theory" into RML research, not only diagnosing current inequities but also proposing feasible methods to advance fairness under real-world conditions.
  2. Differentiates the methodological characteristics of ideal theory and non-ideal practices, while uncovering undocumented implicit steps taken by practitioners.
  3. Provides a systematic framework for explicitly documenting technical interventions and methodological choices.

• Implementation Steps:

  1. Data Issue Diagnosis: Conduct detailed diagnoses of adverse attributes in data that may affect RML goal achievement.
  2. Establish Mapping Relationships:
     • Map data issues to achievable specific goals (Realized Goals, RG) and abstract values (Abstract Values, AV).
     • Document assumptions (Known Assumptions, KA) and interventions (Interventions, IV).
  3. Documentation and Iteration:
     • Update mappings in each development cycle version.
     • Provide dynamic, interactive visualization tools to display relationships between components.

Research Findings

• Key Findings:

  1. Methodological Characteristics:
     • ML practitioners' methodologies exist on a continuum of idealization: some rely on ideal theory (e.g., abstractly addressing RML issues), while others reflect non-ideal methods (e.g., addressing current imperfect conditions).
     • Detailed documentation of non-ideal methods (e.g., diagnosing "non-ideal conditions," integrating values) is insufficient.
  2. Sources of Constraints:
     • Institutional constraints (e.g., limited data access permissions).
     • Motivational constraints (e.g., practitioners lacking intrinsic or external incentives to prioritize fairness-focused work).
     • Resource constraints (e.g., computational and time pressures associated with fairness calculations).
  3. Tool Usage Limitations:
     • Tools used by practitioners (e.g., fairness metrics, explanation tools) often lack comprehensive documentation of uncertainties under non-ideal conditions.

• Experimental Results and Framework Description:

  • A mapping framework comprising five key elements (data issues, specific goals, abstract values, interventions, and assumptions) can achieve clear documentation pathways and facilitate collaboration among multiple stakeholders.

• Comparison with Existing Solutions: This framework emphasizes real-world sensitivity and practical guidance more than existing RML solutions based primarily on ideal theory, while focusing on transparency in technical processes and interdisciplinary collaboration.

• Limitations and Future Directions:

  1. Limitations: Further research is needed to scale the mapping framework for large-scale or complex ML systems, particularly regarding transparency in practices involving incomplete data annotations.
  2. Next Steps:
     • Develop visualization tools to support interactive application of the framework.
     • Extend the framework to address scenarios involving complex non-technical stakeholders or labor-intensive data annotation practices.
     • Investigate how the framework can serve as a public accountability tool, such as "accountability cards."