Considering Process in Algorithmic Bias Detection and Mitigation

Potential for AI to help

• Biased-based on names
• Inconsistent-
• Slow-

algorithms can also bring about bias and inconsistency

• biased against black defendants to be held in jail against white defendants
• need to be careful with these

algorithmic fairness

fairness performance tradeoffs

• accuracy and fairness
• more fair, often less accurate
• doesn't have to be this way

investigates the entire process - her way

real-world interventions, tech policy

TODAY'S TALK

• Leave-one-out unfairness
• IRS collaboration
• Bias in recivid....

Leave-one-out unfairness

• inconsistent
• model changes based on different person being a part of the training set
• this isn't ideal
• the chance that a person's outcome can change on a 1-point change in dataset
• it is unfair, feels arbitrary

learning rule instability as (procedural) unfairness

definition: leave-one-out unfairness

• focuses on consistency
• connections to differential privacy
• how much of a problem is LUF IRL?
  • very prevalent in deep models
• different graphs shown: german credit, seizure, LFW

we need a more formal understanding of inconsistency

how to predict with selective ensembles

consistency and explanations

• inconsistency limits individuals's ability to improve their prediction outcomes
• created method for providing stable

mitigation techniques: ongoing work

• revenue and no change rate