The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Causal Conceptions of Fairness and their Consequences with Sharad Goel - #586
15 snips
Aug 8, 2022 
In this enlightening discussion, Sharad Goel, a Harvard public policy professor and expert on fairness in machine learning, shares his insights on causal conceptions of fairness. He unpacks the limitations of traditional fairness definitions, emphasizing the role of causality in understanding bias in algorithmic decision-making. Goel explores the tension in policy-making between procedural and outcome fairness, revealing how rigid adherence to causal fairness can lead to suboptimal results. He also delves into surprising findings regarding Pareto dominance and its implications for equitable policies.
AI Snips
Chapters
Transcript
Episode notes
Causal Approaches Not a Silver Bullet
- Sharad Goel's prior work highlighted shortcomings in mathematically defining fairness.
- His new work suggests causal approaches, while nuanced, don't solve all fairness issues.
Two Causal Fairness Categories
- Causal fairness definitions fall into two categories: reducing protected trait effects and minimizing counterfactual disparities.
- These consider direct/indirect influence and error rates in alternate decision scenarios.
Difficulty Defining Race's Effect
- Defining "effect of race" is difficult due to the inability to manipulate race directly.
- Studies often use proxies like name changes on CVs, but this simplification requires suspending disbelief.