2022-02-23#

Lead Scribe : Derek

Fairness and Causality/Experiments#

Causality#

Presented by Surbhi

  • Main Goal

    • Teach more how causality can help us understand a situation

  • Passive observation

    • Something we generally observe but it does not require proper attention

    • Like noticing a car driving by you

    • Data collected this way shows a snapshot of the world as it is

  • Causal reasoning

    • Important questions are not usually observational in nature

      • Would traffic fatalities decrease if we raised the legal driving age by 2 years?

  • Not every cause qeustion is easy to address

    • Causal inference gives a formal language to ask questions, but does not trivialize finding the answer

  • Supporting 3 purposes for understanding causality

    1. Conceptualize and address limits of observational techniques

    2. Provide tools to help design intervention that acheive a desired effect

    3. Engage with important debate about when and how much reasoning about discrimination and fairness require causal understanding

  • Simpson’s Paradox

    • Higher acceptance ratio among women, while two show a higher acceptance for men

    • There’s variation between departments, and so the higher rate for men may be seen in aggregate but reversed at the department level

    • This is one example of us misinterpreting what conditional probabilities encode

    • If we look at overall data, it looks like admissions are unfair and discriminating for men, but broken by department it looks like it’s the other way around

      • Not the admission that makes it look this way but it depends on where women/men tend to apply and such

        • For example engineering vs english vs other

  • Causal Models

    • Causal ineference can be used to guide design of studies

      • Choosing which variables to include, which to exclude, which to hold constant

    • Serve as mechanism to incorporate scientific domain knowledge and exchange plausible conclusions

  • Structural Causal Models

    • A sequence of assignments for generating a joint distribution from independent noise models

    • By executing the sequence of assignments, we build a set of jointly distributed random variables

    • Provides a joint distribution and describes how the dist can be generated from elementary noise variables

    • A Probabilistic model

      • Different set of operations

      • Useful for considering hypothetical scenarios differently than if you have a single set of data

      • How we can intervene in a world AND measure the effect of that intervention

  • Causal Graphs

    • Parent notation we saw last week, interpreted a little differently and we can take a model to build a graph

    • Can go from graph to model too

      • Looking at directed graphs as placeholder for an unspecified structuarl causal model that has the assignment structure given by the graph

    • Graph Structure

      • Forks

        • Has outgoing edges to two other variables

        • Aka the most common cause of other variables

      • Mediators

        • A case of a fork where x causes z and z causes y so x causes y multiple ways

      • Colliders

        • Not confounders

        • X and Y are unconfounded, and we can replace do-statements by conditional probabilities

  • The Harvard Admission example

    • The story

      • Asian American Male 25% admission chance, as white 36%, as Hispanic 77%, as African American 95%

    • Invalid statistically because everything on the application is exactly the same except for race but typically many other things will change besides just the race

Empirical Comparison#

Presented by Chan

  • Main Goal

    • Metaanalysis paper of analyzing fair machine learning algorithms

    • Comparison of algorithms they’d found

  • Problem

    • Fairness

      • Comes with sensitive data

      • Age, race, gender, etc

  • The Experiemnts

    • Testing on

      • Preprocessing

        • Comes in 2 ways before feeding into the algorithms that may or may not preprocess further

        • Training data is the cause of discrimination motivates this

        • If training data is discriminating then the results definitely will be, so we can preprocess to make it more fair

      • Modifications to the algorithms

      • Postprocessing

        • If we have a result, can we prune or round items for better results

  • The Results

    • Measures of discrimination correlate with each other

    • Algorhtms make different fairness accuracy tradeoffs

    • Algorithms are gragile: they are sensitive to variations in the input

  • Datasets

    • Ricci

      • Determining if firefighters would receive a promotion

    • Adult Income

      • Predicting income above or below $50k

    • German

      • Classifying people on good or bad credit risk

    • ProPublica Recidivism/Violent Recidivism

      • Committing a crime/violent again

  • Preprocessing

    • Modifying input according to any data-specific needs

      • Removing unneeded features, imputing missing data, etc

      • Add a combined senstivie attribute i.e. “White-Woman”

    • Senstiive Attribute treated as binary

    • Analysis

      • See Figure 2 in the paper

      • One dot is one fold (LOOCV) of the data score in whatever metric in one algorithm

      • Equality between preprocessing choices = Showing the plot is not square, where the x and y axis are equal

      • Binary sensitive attribute is more accurate when using a fair classifier

      • Things we knew about ML before don’t necessarily apply to fair classifiers

        • Fairness may limit our accuracy