Lecture 6 Flashcards

Fairness

1
Q

Fairness: Theory of Fairness & Operational Fairness

What is the difference between equality and equity ?

A

Equality is about equal treatment:
* Fairness of Treatment
* Individual Fairness
* Equal Opportunities
Equity is about equitable treatment:
* Fairness of Outcomes
* Group Fairness
* Creation of Opportunities

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

Fairness (in DA/ML): Legal perspective

What can discrimination be due to?

A

Discrimination can be due to
– Individual practices
– Institutional practices
– structural practices

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

How does discrimination change between dominant and minority groups ?

A

– Dominant: Individuals or groups who have power in society
– Minority: groups that lack power often historically (should be protected)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

What is the theory of justice ?

A

The theory of justice exists through 3 aspects:
* Procedural Justice
* Retributive and Restorative Justice
* Distributive Justice

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

What is procedural justice ?

A

Type of justice concerned with fairness in designing and implementing processes

e.g., delegating power through fair elections …

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

What are Retributive and Restorative Justice ?

A

Retributive and Restorative Justice are retroactive approaches that define the response to past injustice or wrongdoing

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

What is distributive justice ?

A

Distributive Justice is concerned with the fair distribution of resources material (e.g., resources, loans, salaries, housing), services (e.g., healthcare, entertainment, education), or opportunities (e.g., jobs, political participation) and the distribution of burdens (e.g., barriers, duties, taxes, sanctions, and punishments etc

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

What does the resourcist approach of John Rawls say ?

A
  • All people should have equal and equitable levels of resources.
  • Attempts to mitigate inequalities should maximise the benefits of the least advantaged.
    It can be reached through several approaches:
  • Egalitarianism: All groups and individuals are treated equally or equitably.
  • Utilitarianism: The outcomes should be equal or equitable. The desired utility is often linked to the individual’s happiness.

EX: I have a bike, thus I am happy

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

What is a Capabilities Approach to justice

A

The capability approach is a theoretical framework that entails two normative claims:
1. the claim that the freedom to achieve well-being is of primary moral importance,
2. that well-being should be understood in terms of people’s capabilities and functionings.

Ex: I can move around and I can ride a bike. It’s nice

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

At what level is unfairness present in the ML pipeline ?

A

Unfairness is present in the pre-processing as:
1. Historical bias
2. Representation bias
3. Measurement bias
And in the in-processing as:
1. Learning bias
2. Evaluation bias
3. Aggregation bias
4. Deployment bias

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

What is a historical bias ?

A

They occurs when an ML model reproduces or
reinforces a harmful stereotypes already existing in the world.

ex: Scores for white defendants were skewed toward lower-risk categories.
Scores for black defendants were not

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

What is representation bias ?

A

It occurs when the data is not representative of
the population the model is developed for, or specific categories are largely under-represented.

ex: black women identified by AI as white males wearing black shirts

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

What is Measurement bias?

A

They occur when choosing, collecting, or computing the inadequate features and labels to use in a prediction problem.

Features and labels are the concrete measurements chosen to approximate some construct (an idea or concept) that is not directly encoded or observable.

COMPAS ex: number of “arrest” is used measure “crime” or “arrest”

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

What is Aggregation bias ?

A

occur when a one-size-fits-all model is used for
data in which there are underlying groups or types of examples that should be considered differently.

EX: Medical applications often risk aggregation bias e.g., patients of different sexes with similar conditions may present different symptoms and progress in different ways.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

What is Learning bias ?

A

They occur when modelling choices amplify disparities.

EX: prioritising overall accuracy during the learning risks amplifying disparate impact.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

What is Evaluation bias?

A

They occur when the benchmark data, evaluation metrics or evaluation process used to assess a particular task are not adequate or do not represent the use population.

Ex: a robot learning to walk. If we evaluate “walking” as “moving from point A to point B” then this robot is great at “walking”

14
Q

What is Deployment bias ?

A

they occur when there is a mismatch between the model (problem it is intended to solve) and applications (scenarios it is actually used for).
* These biases are of concern, especially in decision-aid-systems.
* Risk Assessment Tools in criminal justice
– Intended to predict the likelihood of committing a future crime.
– Used in o help determine the length of a sentence or early release decisions.

15
Q

What is observational fairness?

A

It’s about equalising observable outputs across groups.
1. Define groups according to
– Sensitive Variables
– Proxies
2. Equalise with the functions from lecture 5
– TP (true positive), TN (true negative), FP, or FN
– Accuracy
– Recall
– …

Usually the features of an individual incorporate all types of biases, sensitive attributes are not known or ill-defined. The classifier mapping features to sensitive attributes is also not well-defined. This results in a poor predicted proxy for the variable of interest as the real outcome if often biased.

16
Q

What is blind fairness?

A

identifying and removing all
sensitive/protect attributes and proxies

17
Q

What are group independent predictions?

A

demographic parity to ensure that predictions are uncorrelated with sensitive attributes (protected attributes and proxies)

18
Q

Equal metrics accross group are…

A

Equal metrics of some sort (e.g., accuracy, true positive rates or false positive rates) across groups

slide 40 (29) of week 6 has an activity but I dont understant it exactly but it might be useful

19
Q

What is causal fairness?

A

It is concerned with causal relations between sensitive attributes and outcomes; these causal relations can be source of unfairness (for obvious and less obvious reasons)

20
Q

How do we mitigate bias and ensure (some) fairness ?

A

Ideally, we would detect bias, mitigate it then assess.
In reality, this is more complex and often has a lot of recurring phases (multiple bias investigations, trade-offs and talks…)