Measuring Algorithmic Bias In Automated Classification Flashcards
How can discrimination in databases be uncovered?
Discrimination in databases can be uncovered by polling people for perceptions of discrimination, studying potential discriminators, and statistical analysis.
What is the Discrimination Discovery Task?
The Discrimination Discovery Task is to find discriminatory situations and practices given a database of historical decision records and a set of potentially discriminated groups.
What are the potentially discriminated or protected groups?
Protected groups are groups defined by socially salient characteristics that experience disadvantages, such as sex, sexual orientation, ethnicity, national origin, age, disability status, and intersections of these groups.
What are the challenges in discovering discrimination in databases?
contextualized non-discrimination requirements,
different conceptualizations of discrimination,
different metrics and criteria,
high dimensionality data,
the possibility of hidden indirect discrimination
What is the relationship between data mining and discrimination?
Data mining can perpetuate discrimination if it is based on biased data or if it amplifies and reinforces existing societal biases.
For example, if a dataset used to make hiring decisions is biased against certain groups of people, such as women or people of color, data mining algorithms trained on that data will learn and perpetuate those biases.
What are the two examples of algorithmic fairness?
The two examples of algorithmic fairness are group fairness and individual fairness.
What is an example method for group fairness?
Classification Rules Mining is an example method for group fairness.
Identifying patterns in data that can be used to create fair decision-making rules that treat different groups equally.
For example, in the context of hiring decisions, Classification Rules Mining could be used to analyze the factors that lead to successful job performance and identify rules that prioritize these factors in the hiring process, while ensuring that these rules are applied equally to all demographic groups.
What are directly discriminatory rules?
Directly discriminatory rules are rules that discriminate against a protected group based on a database of past decisions.
Directly discriminatory rules are rules that explicitly use a protected attribute (such as race, gender, or age) to make a decision or classification about an individual or group, resulting in unfair treatment or exclusion based on that attribute.
What are indirectly discriminatory rules?
Indirectly discriminatory rules are rules that discriminate against a protected group based on an indirect relationship to a protected attribute.
A job requirement for a certain number of years of experience in a particular industry, which may disproportionately affect younger job applicants.
What are genuine occupational requirements?
A genuine occupational requirement (GOR) is a job requirement that is essential for a particular job and is justifiable on non-discriminatory grounds;
for example, only hiring female models to model women’s clothing in a fashion show would be a genuine occupational requirement.
What is the Discrimination Discovery Task limited to?
Limited to identifying the presence of bias and discrimination in machine learning models and does not address the root causes or solutions for these issues.
What is an example method for individual fairness?
We want to use KNN to predict which applicants are likely to be hired by a company, while ensuring individual fairness based on gender.
To achieve this, we can first pre-process the data by creating two separate datasets, one for male applicants and one for female applicants. Then, we can apply KNN separately to each dataset, using a common value for k.
Finally, we can combine the results of the two KNN models to obtain a final prediction that is individually fair with respect to gender.
What is situational testing?
Situational testing is an approach for creating controlled experiments in which matched pairs undergo the same situation.
What is k-NN as situation testing?
k-NN as situation testing is an algorithm that measures the degree of discrimination of the decision for r by looking at its k closest neighbors
Question: What are some predictive tasks that can be performed with given data about a defendant in the criminal justice system?
Answer: Some predictive tasks that can be performed include the probability of general recidivism, probability of violent recidivism, probability of violence against others in prison, probability of self-harm in prison, and probability of breaking permits.
