Fairness Evaluation in Text Classification: Machine Learning Practitioner Perspectives of Individual and Group Fairness

Title of the Paper

Application of Fairness Assessment in Text Classification: Perspectives of Machine Learning Practitioners on Individual and Group Fairness

Paper Information

• Domain: Fairness assessment in machine learning and natural language processing
• Keywords: Fairness assessment, text classification, group fairness, individual fairness, human-computer interaction, toxic text detection, algorithmic bias, transparency

Research Background and Problem Statement

• Issues and Challenges:

  • Machine learning models may lead to social inequities due to algorithmic bias, such as inaccuracies in skin color recognition, representational bias in search engines, and gender bias in translation services.
  • Although various toolkits have been developed to help practitioners address algorithmic bias (e.g., fairness assessment tools), there is a lack of in-depth research on how model fairness is evaluated in practice.
  • User exploration of different dimensions of fairness (e.g., individual fairness and group fairness) in text classification remains insufficiently investigated.

• Significance of the Study:

  • Ensuring fairness in artificial intelligence not only has a positive impact on societal development but also often involves compliance with legal and regulatory requirements.
  • Reducing unfairness in NLP systems toward different demographic groups contributes to improving the ethicality and trustworthiness of AI technologies.

• Motivation for the Study:

  • Address the current gap in understanding machine learning practitioners' perspectives on algorithmic fairness assessment.
  • Investigate users' exploration strategies and underlying motivations when interacting with fairness tools in text classification.

Proposed Solution

• Methodology:

  • Develop an interactive tool that allows users to dynamically explore the fairness of text classification models, including:
    • Group fairness metrics
    • Individual fairness metrics
    • A custom input experimentation mechanism enabling users to explore model performance on specified inputs.

• Innovations:

  • Investigate how users' decisions regarding group or individual fairness dimensions are influenced by diverse but static metrics.
  • Provide grouped data views and interactive single-instance views to study how practitioners customize inputs and favor different experimental approaches.

• Implementation Steps and Key Techniques:

  1. Build an interactive tool for fairness exploration, supporting four modes:
     • Predefined group fairness view
     • Predefined individual fairness view
     • Custom group input view
     • Custom individual input view
  2. Use contextualized experiments to showcase the performance and fairness data of three different text classification models (standard model, individual fairness model, group fairness model).
  3. Collect user model selection data and strategies under each type of metric.
  4. Provide comprehensive feedback on individual experiments corresponding to fairness dimensions, including trends in metric changes.
  5. Analyze participants' preferences for model selection based on interaction and fairness metrics, distilling subjective preferences for fairness indicators.

Research Findings

• Specific Outcomes:

  • Developed a front-end interactive tool for exploring group and individual fairness in text classification tasks.
  • Key findings include:
    1. The types of fairness metrics provided (e.g., accuracy and fairness) significantly influence users' judgments on which model is the "most fair."
    2. In custom input experiments, users tend to focus on input dimensions related to their personal experiences or current societal issues.
    3. Users adopt different strategies when exploring group fairness and individual fairness, such as grouping by similar identity categories or attempting to generate "adversarial inputs."

• Advantages:

  • Experiments demonstrate that presenting multi-perspective fairness evaluations (e.g., group and individual fairness) helps users make fairness assessments that better align with real-world needs compared to single evaluation criteria.
  • The tool design exhibits strong user flexibility, allowing practitioners to deeply customize and analyze potential model biases.

• Experimental or Evaluation Results:

  • In Task 1 (focused solely on prediction accuracy), most users selected the standard model (88%). However, in Task 2 (group fairness view), 50% of users switched to the group fairness model. In Task 3, 67% of users chose the individual fairness model.
  • Custom experiment outputs showed that users' understanding and preferences for fairness indices largely depend on the types of views and the flexibility of interactions provided by the tool.
  • Users preferred models that maintained accuracy while avoiding unfair predictions for specific groups.

• Limitations and Future Directions:

  • The study primarily focuses on specific tasks (e.g., text classification), and further validation is needed to assess the tool's applicability in other domains (e.g., recommendation systems or speech recognition).
  • Using lexical matching methods to filter minority group data may overlook implicit content associations (e.g., toxic content not explicitly mentioning minority groups).
  • Future research could optimize individual and group fairness guidelines and explore ways to provide customized interactive experiences for different user groups.

Through this study, the authors recommend enhancing fairness assessment tools by incorporating contextual information, designing inter-group similarity prompts, and optimizing asymmetric counterfactual metric generation to improve the tools' practicality and effectiveness.