Algorithmic Bias Machine Learning An area of AI that studies how to - - PDF document

Algorithmic Bias

Machine Learning  An area of AI that studies how to get computers to learn from experience (e.g. data)  Identify patterns from a training dataset  Then generalize from these patterns and apply it to future data (that is different). This is called a test dataset  Supervised learning

• Features -> Classifier -> Class Label
• Features: traits of a data instance (e.g. keywords in

your email) that are informative as to the classification

• Class Label: the classification (e.g. Personal or School

for email sorting)

• Produce the classifier by training on the training data

Algorithmic Bias  What is it?

• Bias introduced to machine learning due to the training

data

• Garbage in / Garbage out: machine learning algorithms

reflect societal bias when applied to biased data (bias in the form of discrimination, prejudice and unfairness)

• Do machine learning algorithms respect protected

variables?

• These are characteristics that anti-discrimination laws

protect in certain situations

• E.g. Fair housing act prevents landlords from

discrimination based on 7 protected classes:  Race

 Gender  Religion  Disability  Color  National Origin  Family status

• Can’t just ignore features that correspond to these

protected variables and say your algorithm is not biased  Due to confounding factors e.g. Zip code and race are closely correlated in many parts of the US

• What are causes of algorithmic bias?
• Biased training data (e.g. biased class labels)
• Inclusion of protected variables as features;

inclusion of variables correlated with protected variables are highly problematic

• Downstream goals (e.g. business profitability)

might conflict with discrimination

• Misunderstanding / misuse of machine learning

 Machine learning applied to the wrong tasks

• Domain adaptation: machine learning algorithm

trained on data from one distribution but applied to test data from another distribution

• Missing / corrupted data
• Sampling selection bias
• How do you fix this problem?
• Not sure if you can:

 A lot of these are societal problems

 Can you correct the bias without introducing bias of a different sort?

• Understand the problem so that you can use the

right machine learning algorithm  Know when NOT to use a particular algorithm

• Make systems that are auditable.
• Have less high-impact outcomes earlier on,

especially when an algorithm is involved

• Difficult problems that make a solution hard
• Limited to what data you actually have. What

about the data you don’t have?

• Definitions of fairness vary greatly – which one do

you use?

• Lack of social context: can’t transfer a machine

learning algorithm from one context to another