Ctrl+K

Fairness

Fairness#

Lead Scribe: Chan

FAIRNESS AND MACHINE LEARNING#

Chapter 2: Classification#

Presented by Damon

Main Goal#

  • determine plausible value for an unknown varible Y given an obscent variable X

  • Essentially trying to predict something based on other factors

Supervised learning#

  • The prevalent method for onstructing classifeirs from observed data

    • Goal: to identify menatingful patterns

  • Classifier: a mapping from the space of possible values for X to the space of values that the target Y can assume.

  • The essential idea is simple in that we will have albeld ata in the form (x1, y1) that is drawn from a distribtuion where x1 is an instance and y1 is the label.

    • The instance are then partitioned into positive

    • Lables typically come from a discrete set such as (-1, 1) in the case of binary classification.

Statistical Classification Criteria#

  • How to identify which classifer is best fo your purpose

  • Accuracy: the probability of correctly predicting the target variable

  • Cond. probability table

    • The true positivie rate corresponds to the frequency of the classifer correctly assinging a positive label to a positive instance.

  • Why do we need the weighted average?

    • P(Y=Ŷ) = P(Ŷ=1|Y=1)P(Y=1) + P(Ŷ=0|Y=0)P(Y=0)

    • to get all of the accuracy, we need to multiple the weights to our conditional probabilities so that it is computable.

The Conditional Expectation#

  • Score: a single real-valued variable summairzed from a regression model

  • A natural score function is the expectation of the target variable Y conditional on the features X we have observed

Sensitive Characteristics (A)#

  • In many classification tasks, the features X contain or implicitly encode sensitive characteristics of an individual

  • It is dangerous to ignore these factors, which will make prediction difficult by tampering the correlation.

Indepdence – R⏊A#

  • Requires the sentitive charactieric to be statistically independent of the score

  • Definition 1: The random variable (A, R) satisfy independence if A⏊R

  • Independence simplifies to the condition:

    • P{R=1∣A=a}=P{R=1∣A=b}

  • Where R=1 is acceptance and the condition requires the acceptance rate to be the same in all groups.

  • There coudl also be a relaxation on the constraint that would introduce a positive amount of slack ε>- and require that:

    • P{R=1∣A=a}≥P{R=1∣A=b} − ϵ

    • = ϵ ≥ P{R=1∣A=b} - P{R=1∣A=a}, where we want the difference to be small.

Separation#

  • Acknowledges that in many scenarios, the sensitive characteristic may be correlated with the target variable.

  • This separation criterion allows corelation between the score and the senstitive attribute to the extent that it is justified by the target variable

  • Definition 2: Random variables (R,A,Y) satisfy separation if R⏊A|Y

  • In the case where R is a binary classifier, separation is equivalent to requiring for all groups a, b the two constraints:

    • P{R=1∣Y=1,A=a}=P{R=1∣Y=1,A=b}

    • P{R=1∣Y=0,A=a}=P{R=1∣Y=0,A=b}

  • Separation requires that all groups experience the same false negative rate and the same false positive rate.

Sufficiency – Y⏊A|R#

  • Formalizes that the score already includes the sensitvie characteristic for the purpose of predicting the target

  • Definition 3: We say that random variables (R, A, Y) satistfy sufficienty if Y⏊A|R

  • In this case, a random variable R is sufficient for A iff for all groups a, b and all values r in the support of R, we have

  • When R only has 2 values we recognize this condition as requiring a parity of pos./neg. predictive values across all groups.

Calibration and Sufficiency#

  • It is sometimes desirable to be able to interpret the values of the score functions as probabilities

  • This condition means that the set of all instances assigned a score value r has an r fraction of positive instances among them.

Calibration by Group#

  • To formalize the connection between sufficiency and calibration we say that the score R satisfies calibration by group if it satisfies:

  • Proposition 1. If a score R satisfies sufficiency, then there exists a function 𝓵: [0,1] --> [0,1] so that 𝓵(R) satisfies calibration by group.

Relationships Between Criteria#

  • Initial criteria constrains the joint distribution in non-trivial ways therefore we can suspect that imposing any 2 of them simultaneously over-contrains the space to the point where only degenerate solutions remain.

  • Happy path of establishing multiple fairnesses is intangible.

Machine Bias#

Presented by Derek

The premise#

  • Brisha Borden & a friend - 18 years old

    • Scooter

  • Vernon Prater - “Seasoned”

    • Shoplifting

  • Both similar crimes

  • Computer program predicting likelihood of recidivism

    • Borden > Prater

    • Two years later, opposite held true

Risk Assessments#

  • Becoming increasingly more popular in courtrooms

  • 9 states provide scores to judges during sentencing

Following the Warnings of AG Holder#

  • The sentencing commission did not launch a study of risk scores

  • ProPublica did launch a study

    • >7000 people arrested in Broward Cty, FL

    • Check how many were charged with new crimes over 2 years

  • Results

    • Unreliable

    • Biased

In Theory#

  • Risk scores are great

  • High scores mean more likely to commit crime

  • Vice versa for low scores

  • Simple…?

In Practice#

  • People are extremely complex

  • Countless factors go into recidivism such as

    • employment

    • housing status

    • fiancial situations

  • As a result unless we individualize the risk assessment, result will be biased

The Problem#

  • people tend to blindly trust tech.

  • any jurisdictions adopt Northpointe’s software without testing

  • Using tech is the easy way out

    • “Easy to use and gives ‘simple/effective’ charts for judicial review”

  • 2009 study reported a 68% accruacy rate for the scoring software

Overall#

  • Using tech in social context is extremely complex

  • Poeple are unique and countless factors go into individual behavior

  • With recidivism specificialy

    • Black people more likely to be predicted higher than White people

    • Software relied upon too much

    • Studies have concluded risk scores do not reflect reality.

Gender Shades#

Presented by Derek

Intro#

  • AI is widely used in society

  • Facial recognitiion software can realistically be used to identify suspects

  • Algo. trained with biased data result in algorithmic discrimination

  • This work focuses on facial recognition to gender classification

Intersectional Benchmark#

  • Gender classification means we need defined classes

  • The dataset should contain varied phsyical attributes for subgroup accuracy analysis

  • Pilot Parliaments Benchmarks dataset

    • high quality photos

    • reliable sources

Commercial Gender Classification Audit#

  • Classifiers perform best on lighter, male faces

  • Classifiers perform worst on darker female faces

    • Microsoft, IBM, Face++ analyzed

  • Max disparity in error rate between best/worst classified groups is 34.4%

Dissecting racial bias in an algorithm used to manage the health of populations#

Presented by Derek

  • Black people had to have more chronic conditions to have a higher risk score

  • Training used total medical expenditure.

    • Could have used total sickness as a factor to train to avoid bias.

Contents