Cynthia Dwork: Recent Developments in Algorithmic Fairness

Part of my series of notes from ICLR 2019 in New Orleans.

Introduction

• algorithms are unfair
  • unrepresentative / biased training data
  • feature choice can be biased
• …but by now, we all know this, so what do we do?
• many natural desiderata of fairness are proven to conflict
• want a theory of algorithmic fairness
• two kinds of fairness to think about
  • group fairness
    • e.g. statistical parity, balance for positive
    • easier but fail under scrutiny
  • individual fairness
    • people similar wrt task should be treated similarly
    • strong legal framework
    • but how do we define this task-specific similarity metric?

• metric learning for individual fairness
  • distances for single representative element are useful but limited
  • use small number of additional representatives to help
  • this is work by Christina Ilvento but I cannot for the life of me find a paper (I think it might be too recent?)
• “multi-x” approaches as a way to combine group and individual fairness
  • apply group fairness, but simultaneously across all pairs of sets in some collection

Fair Scoring and Ranking

• scores as individual probabilities?
  • e.g. probability of recidivism
  • what does this mean when the experiment isn’t repeated?
• “On Individual Risk” – Dawid 2018
• one way to handle this is calibration
  • e.g. in forecasting – 70% of days where predict 70% chance of rain should have rain
• how to choose groups to compare
  • complexity theory! – all groups identified by small circuits in data
  • data needs to be rich enough – differentially expressive
• multi-accuracy (expectations of intersections etc. work out) and multi-calibration
  • this is computationally doable, wow
• paper on multi-calibration
• the devil is in the choice of collection (they think they have a test for this…)

Fair Representation

• remove sensitive info (censoring) while retaining sufficient info for training
• adversarial learning of fair representations
  • encoder tries to hide group membership
  • decoder tries to reconstruct
  • predictor tries to classify (based on encoded representation)
  • adversary tries to determine group membership
  • papers: Edwards & Storkey 2016, Madras et al. 2018
• can apply same methods to try and find common features among different populations
  • e.g. patients at different hospitals