Link Search Menu Expand Document

Algorithmic Fairness

Algorithmic Fairness from a Non-ideal Perspective

  • important moments:

  • Philosophy of justice

    • https://plato.stanford.edu/entries/justice/
      • obligation ("one's due") in contrast to charity
      • resolve conflicts when interests clash
      • impartiality - two cases relvantly alike should be treated similarly
    • convservative (conserving existing norms) vs ideal (demand reform of norms and practicies)
    • corrective (bilateral, wrong-doer <> wronged ) vs distributive (allocating goods to individuals, mutltilateral with a distributing agent)
    • procedural (procedure) vs substantive (end result)
    • comparative (a position was offered to a less qualified candidate) vs non-comparative (regardless of comparison, e.g. basic human rights)
      • ML disccussion often focus on comparative. should also consider non-comparative
    • scope of justice
    • ideal and non-ideal theorizing (The Imperative of Integration - Elizabeth S. Anderson)
      • ideal:
        • imagine perfectly just world (Rawls)
        • try to minimzie discrepancy between reality and ideal
        • e.g. been used to argue against affirmative acction - ideal world is color blind
      • non-ideal:
        • understand causual explanation of the problem, determine waht can be done and who to correct it.

  • economics

    • "the economics of discrimination" - Becker
      • simulation of employer taste-based discrimination
    • Arrow's Rebuttal to Becker 1973
      • imperfect information as alternative cause
    • "Statistical theory of racism and sexism" 
         - Efficient candidate screening under multiple tests and implications for fairness
      https://arxiv.org/abs/1905.11361 
             - concerns noise 
       - by adjusting number of interviews. hitting boundary on left, can equalize FP and FN across groups w different noise level 
           - but more num of interviews can negatively affect candidates too 
   - Will Affirmative-Action Policies Eliminate Negative Stereotypes?
       - Stephen Coate and Glenn C. Loury
       - https://www.brown.edu/Departments/Economics/Faculty/Glenn_Loury/louryhomepage/papers/Coate%20and%20Loury%20(AER%201993).pdf
       - even when groups are equal ex ante, equilibrium outcomes following some internventions can appear to confirm negative sterotypes

  • ML fairness

    • https://www.theatlantic.com/technology/archive/2018/05/machine-learning-is-stuck-on-asking-why/560675/?utm_source=twb

      • "ML is stuck on ... learning associations "
      • it learns associations and not causal relations
      • on one hand, curve fitting turned out useful many places
      • but in many problems, curve fitting is not enough

    • http://approximatelycorrect.com/2016/11/07/the-foundations-of-algorithmic-bias/

    • taking inspiration from law:

      • title 7 of civil rights law
        • disparate treatment
          • addresses intentional discrimination
            • protected characterstic
            • also via proxy variables
            • with exceptions e.g. if goal is to promote diversity
          • what does intention mean in ML context?
        • disparate treament
          • 3 tests:
            • plaintiff must demo statistical disparity (4/5 rule)
            • defendent must show descirions are justified by business necessity
            • plaintiff must show defendent can achieve goal w alternative practice
          • first one can be done using stats.
          • later 2 of the 3 tests of the above are not well adressed by ML, require causal reasoning

  • typical

    • treatment parity
      • output does not depend on sensitive characteristic
        • Model cannot use that feature
    • impact parity
      • outcome independent of group status
        • model can use the feature, but result algo outcome is indenpdent
    • representational parity
      • the input map to some representation that you can't infer their demographics
      • entails impact parity
    • equalized odds / opportunity parity
      • equal FN and/or FP rates

  • problem

    • the different parities are mutually irreconcilable
    • statisctical parity may not capture legal/philosophical notions
    • lack ingredients to determine just action
      • how did disparities arive
      • impact of the decision
      • responsibilities of the decision maker

  • "impossibility" theorem:

    • if we start from a non-ideal world, no set of action can simultaneously satisfy all the ideal
    • meeting the ideal in some respect may require widening other gaps
    • "equity" - peyton young, 1994 
      - different definition of equity which all seem reasonable by itself, when together causes non-reconcible conflict
      => must make some choices

  • problem applications of attempts for fair algorithm

      • for maximizing impact disparity, treatment disparity is optimal (theortical)

      • if other features sufficiently can encode the sensitive feature, result is indistinguishable from teatment disparity (theortical)

      • if other features partially encode sensitive feature => empirical side effects

        • recorders within group that makes no (not procedurally justifiable)
        • produces potentially bizarre incentives to conform to steortype

      • example case study of gender study in CS admissions

        • when applied DLP, the decisions were flipped neg to pos, for candidates that based on other traits were more likely to be female and vice versa
          • in effect hurt female candidates who were applying to fields that are more male dominated

  • interesting research to follow

    • causal approaches to fairness
      • counterfactual fairness (Kusner 2017)
      • causal explanation (bareinboim 2017)
      • sensitive to subjectivity (different interpretations of the cause)
      • outsource the key issue to humans
    • feedback loops - next-step or equilibrium outcomes in a dynamic model
      • Delayed impact of fair ML, Liu et al
      • Social Cost of strategic classification / Disparate effects of Strategic manipulation
      • Runaway Feedback Loops in predictive Policing