From Fitting Participation to Forging Relationships: The Art of Participatory ML

Title of the Paper

From "Adapting Participation" to "Building Relationships": The Art of Participatory Machine Learning

Paper Information

• Subject Areas: Human-Computer Interaction (HCI), Machine Learning (ML), AI Ethics
• Keywords: Participatory methods, machine learning, artificial intelligence, design, ethics, algorithmic fairness, data subjects, domain knowledge, AI education, social impact

Research Background and Problems

• Identified Issues or Challenges:

  1. The development of Participatory Machine Learning (Participatory ML) reveals that, despite its goal of involving end users and affected individuals in the design and development process, there are practical challenges such as power imbalances and informational conflicts.
  2. Current approaches face the risk of "participation-washing," where participation is used symbolically rather than as a meaningful collaboration.
  3. Balancing the complex contextual information generated during participation with the structured data formats required for machine learning development remains a challenge.

• Importance: Machine learning systems often perform poorly for marginalized groups, and issues of algorithmic bias and system unfairness urgently need to be addressed. Participatory ML can help tackle these problems through democratized design and transparent development processes.

• Research Motivation and Related Work:

  1. This study focuses on researchers and practitioners who act as "participation brokers," exploring how they facilitate inclusivity and manage power dynamics in participatory ML projects.
  2. It aims to expand the scope of Participatory ML research to include projects outside North America and Europe, which are underrepresented in existing literature.

Proposed Solutions

• Methods or Solutions Proposed: The authors investigate the role of participation brokers through 18 in-depth interviews, exploring strategies and challenges in participatory ML projects. This approach seeks to understand how "meaningful participation" can be achieved across diverse organizational and geographical contexts.

• Innovations:

  1. Introduced a new perspective on "participation brokers," focusing on how these individuals balance value creation between design teams and participants through practical actions.
  2. Advocated for a transformation of the broker role, shifting from mere participation facilitators to educators and advocates for action.

• Implementation Steps and Key Techniques:

  1. Interview Design: Recruitment of participation brokers through purposive and snowball sampling for semi-structured interviews.
  2. Data Analysis: Reflexive thematic analysis was used for coding and theme extraction.
  3. Methodological Improvement: Included projects outside North America and Europe, particularly from the Global South, to broaden research perspectives.

Research Outcomes

• Specific Findings:

  1. Described the diverse forms of "participation" in participatory projects and revealed how brokers manage power dynamics.
  2. Proposed tools and structured methods for educating end users, helping them understand how machine learning works and its applications.
  3. Extended participatory methods to address the needs and feedback of indirect stakeholders (e.g., groups affected by algorithms but not direct users).

• Advantages and Comparisons:

  • Compared to existing methods, this study emphasizes viewing participants as collaborators shaping the future of AI, rather than merely data providers or passive participants.
  • Provides specific recommendations for connecting complex systems, including addressing the needs of indirect stakeholders.

• Experimental and Evaluation Results:

  1. During the model evaluation phase, participant feedback was used as a key reference for determining system readiness for deployment.
  2. Brokers found that while data collection and annotation stages were easier to incorporate participation, decision-making power remained concentrated among brokers and data scientists.

• Limitations and Future Directions:

  1. Research Limitations:
     • Small sample size, limiting the ability to fully capture the diversity of participatory methods across industries or countries.
     • Did not directly interview participants in participatory projects, relying instead on brokers' perspectives, which may lead to partial information loss.
  2. Future Directions:
     • Develop tools and frameworks to support participants in expressing complex contextual information, expanding the applicability of participatory ML.
     • Further study mechanisms for translating feedback from indirect stakeholders, especially in projects based in the Global South.

Conclusion and Insights

The future development of participatory machine learning has moved beyond merely adapting existing engineering processes to focus on building long-term relationships, educating end users, and listening to indirect stakeholders. This approach not only enhances the social adaptability of systems but also advances the practice of algorithmic fairness and ethics. The authors recommend that participation brokers evolve into educators and advocates to unlock the democratic potential of participatory machine learning.