Session 7 Flashcards
Sensitive Characteristics (or protected attributes)
are those that cannot be used (legally) to differentiate individuals with respect to the target variable in predictive models
By defining some characteristics as “sensitive”, we are assuming that the algorithms can end up differentiating individuals based on these characteristics
- This may be because the data reveals existing injustices (e.g., some groups of individuals may be already discriminated, and that shows in the data)
- It can be because of differences in tastes and behaviors, and may not represent discrimination (e.g., Sinterklaas is more popular in The Netherlands than in Portugal)
Formal non-discrimination criteria
Many fairness criteria have been proposed over the years, each aiming to formalize different requirements
Most proposed fairness criteria are properties of the joint distribution of:
A - the sensitive attribute
Y - the target variable
R - the classifier or score
Most criteria fall into one of three categories regarding how these variables are related with each other:
- Independence (R ⊥ A)
- Separation (R ⊥ A | Y)
- Sufficiency (Y ⊥ A | R)
Independence
has been explored through many equivalent terms or variants, referred to as demographic parity, statistical parity, group fairness
Main idea: “Everybody gets treated the same”
A classifier R is independent from an attribute A if
the probability of the classifier predicting an observation to be positive (R = 1) does not change with a change in the attribute A:
Pr(R = 1 | A = a) = Pr(R = 1 | A = b)
Example: The probability that a person is predicted to default on their loan does not depend on their race
Seperation
R ⊥ A | Y
requires the score (R) to be independent from the sensitive attribute (A) given the outcome (Y). In other words, it allows correlation between the score and the sensitive attribute to the extent that it is justified by the target variable
Seperation (R ⊥ A | Y)
Main idea:
Given an outcome (e.g., defaulting on a loan), the percentage of individuals predicted positive (and negative) are similar across groups of a sensitive attribute (e.g., black, white)
Sufficiency (Y ⊥ A | R)
requires the outcome (Y) to be independent from the sensitive attribute (A) given the score (R). In other words, it allows correlation between the outcome and the sensitive attribute to the extent that it is justified by the score.
Sufficiency (Y ⊥ A | R)
Main idea:
Given a prediction, the percentage of those that are positive is similar across groups of a sensitive attribute (e.g., black, white)
PPV / NPV
Positive Predicted Value
Negative Predicted Value
Relationships between criteria
In general
Each of these fairness criteria is incompatible with the other two. You can satisfy only one of them at a time.
How is Google able to find cats in my photos?
Approach #1: Predictive modeling
- Define a target variable
- > cat vs no cat - Gather a large set of photos
- > label the photos - Create a set of features (or predictors)
- > 2 eyes, pointy ears, spots - Run a tree induction model
- Use the model to classify my photos
Deep learning
is a new area of machine learning that uses artificial neural networks for unsupervised pattern recognition
