Explaining models: an empirical study of how explanations impact fairness judgment
summary
Study builds two logistic classifiers from Broward County data (raw and debiased via a preprocessing step), generates ~1,000 programmatic textual explanations of four explanation styles, and surveys >160 Mechanical Turk workers who each rate six cases (two disparate impact cases over‑sampled, four non‑impact). Processed classifier is rated significantly fairer than raw; local explanations (sensitivity, case‑based) reveal disparate impact more effectively than global ones; higher trust in ML raises fairness ratings. Recommendation: combine global and local, personalize explanations.
AI-written summary · about
chapters (10)
- 0:00Introduction and motivation
- 1:29Study overview and methodology
- 2:55Data source and debiasing process
- 4:21Explanation types and their generation
- 7:12Disparate impact and sampling
- 7:55Survey design and participant factors
- 8:18Key results on perceived fairness
- 11:31Geometric interpretation of explanations
- 12:37Key takeaways and lessons learned
- 14:10Q&A session
related talks
20m`It's Reducing a Human Being to a Percentage'; Perceptions of Justice in Algorithmic Decisions
34mInterpretability and Algorithmic Fairness
28mBeyond Prediction Performance: How Modeling Decisions Shape Fairness Outcomes in Statistical...
39mGuide towards algorithm explainability in machine learning
48mInherent Trade-offs with the Local Explanations Paradigm
31mMeasuring Model Fairness
46mAI for High-Stakes Decision Making
46mA look at the methods to detect and try to remove bias in ML models