Indirect Discrimination and Algorithmic Fairness, AAAI Fall Symposium 2020 (AI for Social Good)

The research explores potential gender bias in gambling harm detection algorithms and proposes frameworks for algorithmic fairness.

Categories

Chris Percy, representing Playtech and City of London, presents research on gender bias in machine learning models used to identify problem gambling. The study explores two key concerns: indirect discrimination and algorithmic fairness, with a focus on public health applications in the UK gambling sector.

Key Insights:

• Indirect Discrimination Test:
  A novel method—model-matched indirect identification—was used to assess whether gender could be inferred indirectly by the model. Results showed no significant indirect discrimination across two operators.
• Gendered Performance Disparities:
  By incorporating voluntarily provided gender data, researchers uncovered performance differences—particularly in a female-dominated bingo brand, where the model was more accurate for identifying at-risk female players.
• Rejected Approach – Direct Gender Input:
  Including gender as a model feature did not improve performance and slightly worsened outcomes, especially for female players. The feature ranked low in importance.
• Preferred Approach – Blind Separate Models:
  Separate models were trained for male, female, and unclassified players. Each player was run through all models, and the highest risk score was used. This reduced gender disparity in risk identification, especially improving male detection, but increased false positives and reduced overall accuracy.
• Limitations & Ethical Considerations:
  Gender and harm definitions were simplistic, and fairness remains a contested concept. The team advocates for sector-specific ethical AI frameworks and cross-industry learning to improve responsible gambling tools.

This presentation highlights the complexity of fairness in AI-driven gambling harm detection and the need for nuanced, stakeholder-informed approaches. For further discussion, viewers are encouraged to engage with the research team or join sector working groups.