Keynote - Fairness and Diversity in Online Social Systems
summary
Algorithms systematically amplify real-world biases, shown by COMPAS recidivism, search-order effects on voting, skewed Google Images and biased ad delivery. Concrete tools include constrained multi-armed bandits that enforce box-like upper and lower bounds on group exposure and a fast greedy ranking algorithm that respects top-K group caps. Empirical claims: low-regret personalization with a small or often negligible price of fairness, and diversity in forum rankings improves learner outcomes rather than degrading them.
AI-written summary · about
chapters (11)
- 0:00Fairness in algorithmic systems
- 2:13Algorithmic bias examples
- 7:29Defining fairness in decisions
- 10:20Breaking algorithmic bias
- 17:28Group fairness framework
- 26:24Fairness in content ranking
- 30:56Reusing fairness algorithms
- 36:34Fairness impact on users
- 43:38Fairness metrics debate
- 45:03Fairness research challenges
- 50:02Summary & Q&A
related talks
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
35mOpen Source Fairness
53mGroups and Individuals in Fairness and Probability
42mFairness in Machine Learning: From Theory to Practice
11mAddressing Privacy, Bias, and Fairness in AI Systems
46mEliciting and Enforcing Subjective Individual Fairness
38mSocial Implications of Bias in Machine Learning
31mAvoiding Digital Bias