This module introduces a framework for approaching fairness in machine learning, including defining and checking for fairness and considerations when choosing between different fairness implementations.
Fairness Criteria slides (PDF - 1.5MB)
In the case of a binary classification (for example, whether or not to hire someone), you can categorize values in four categories. TP = true positive (correctly classified as positive). TN = true negative (correctly classified as negative). FP = false positive (incorrectly classified as positive). FN = false negative (incorrectly classified as negative).
The outcome is independent of the protected attribute. For example, the probability of being hired is independent of gender. Demographic parity almost always cannot be implemented if individuals are members of multiple protected groups because you may not be able to impose the equal probabilities across all groups. Demographic parity can also be fair at a group level, but unfair at an individual level. For example, if qualifications are different across a protected attribute, imposing demographic parity may mean someone who is less qualified may get hired. Therefore, if a large number of unqualified male applicants is added to the applicant pool, the hiring of qualified female applicants will go down.
Equalizing the odds means matching the true positive rates and false positive rates for different values of the protected attribute. This means that we are only enforcing equality among individuals who reach similar outcomes. This algorithm is more challenging to implement, but achieves one of the highest levels of algorithmic fairness. For example, the probability for a qualified applicant being hired and the probability of an unqualified applicant not being hired should be the same across all protected attributes. As compared to demographic parity, if a large number of unqualified male applicants apply for the job, the hiring of qualified female applicants in other protected groups is not affected.
Equalized opportunity means matching the true positive rates for different values of the protected attribute. This is a less interventionist approach of equalizing the odds and may be more achievable. In the example of hiring, for qualified applicants, the algorithm would work exactly as the equalized odds algorithm. For unqualified applicants, the rates of not hiring would not be same across different values of the protected attributes. For example, unqualified men would not necessarily have the same rate of not being hired as unqualified women.
Hardt, Moritz, Eric Price, and Nati Srebro. “Equality of opportunity in supervised learning.” Advances in Neural Information Processing Systems. 2016.
Kilbertus, Niki, et al. “Avoiding discrimination through causal reasoning.” Advances in Neural Information Processing Systems. 2017.
Wadsworth, Christina, Francesca Vera, and Chris Piech. “Achieving fairness through adversarial learning: an application to recidivism prediction.” arXiv preprint arXiv:1807.00199 (2018).
Pleiss, Geoff, et al. “On fairness and calibration.” Advances in Neural Information Processing Systems. 2017.
Verma, Sahil, and Julia Rubin. “Fairness definitions explained.” 2018 IEEE/ACM International Workshop on Software Fairness (FairWare). IEEE, 2018.
Content presented by Mike Teodorescu (MIT/Boston College).
This content was developed in collaboration with Lily Morse and Gerald Kane (Boston College) and Yazeed Awwad (MIT).
Protected Attributes and “Fairness through Unawareness” slides (PDF)
Disparate treatment is when you are disproportionately favoring a particular protected class by intentionally including variables tied to protected attributes. Disparate impact is when you are disproportionately favoring a particular group unintentionally.
Protected attributes are features that may not be used as the basis for decisions. Protected attributes could be chosen because of legal mandates or because of organizational values. Some common protected attributes include race, religion, national origin, gender, marital status, age, and socioeconomic status.
When implementing a machine learning algorithm, you start by collecting data. Data contains outcome variables as well as predictor variables. The complete dataset is split into a training set where you run your model and a test set where you test your model to determine model performance. Implementing fairness starts with the training set. The training set can carry the biases of the people labeling data. In the case of resumes, if you have a set of graders and those graders have biases, those biases could be reflected in how the data is labeled. Bad training data will lead to a bad prediction. The training data may not be representative of all groups (for example, gender). Past data may not predict current events. Risks from training data include biased sampling, small sample sizes for certain subgroups, hidden correlations, the possibility that the data may contain protected variables, and a large degree of noise.
Fairness through unawareness assumes that if we are unaware of protected attributes while making decisions, our decisions will be fair. This approach has been shown to not be effective in many cases. Protected variables could be correlated with other variables in the data. For example, race or religion may be linked to a city or neighborhood in a city.
Abdi, H. “The Kendall Rank Correlation Coefficient.” Encyclopedia of Measurement and Statistics. Sage, Thousand Oaks, CA (2007).
Agrawal, Ajay, Joshua Gans, and Avi Goldfarb. Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Press, 2018.
Ajunwa, Ifeoma. “The Paradox of Automation as Anti-Bias Intervention.” [Forthcoming in Cardozo Law Review (2020).] March 10 (2020): 2016.
Angst, Corey M., and Ritu Agarwal. “Adoption of electronic health records in the presence of privacy concerns: The elaboration likelihood model and individual persuasion.” MIS Quarterly 33.2 (2009): 339–370.
Angst, Corey M. “Protect my privacy or support the common-good? Ethical questions about electronic health information exchanges.” Journal of Business Ethics 90.2 (2009): 169–178.
Apfelbaum, Evan P., et al. “In blind pursuit of racial equality?” Psychological Science 21.11 (2010): 1587–1592.
Ashcraft, Catherine, Brad McLain, and Elizabeth Eger. _Women in Tech: The Fact_s. National Center for Women & Technology (NCWIT), 2016.
Content presented by Mike Teodorescu (MIT/Boston College).
This content was developed in collaboration with Lily Morse and Gerald Kane (Boston College) and Yazeed Awwad (MIT).
