Bias and Fairness

Today’s post has AI and Policy news updates and our next installment on Bias and Policy: the fairness component.

News Items for February, 2020

  • OECD launched the OECD.AI Observatory, an online platform to shape and share AI policies across the globe. 
  • The White House released the American Artificial Intelligence Initiative:Year One Annual Report and supported the OECD policy.

Bias and Fairness

In terms of decision-making and policy, fairness can be defined as “the absence of any prejudice or favoritism towards an individual or a group based on their inherent or acquired characteristics”.  Six of the most used definitions are equalized odds, equal opportunity, demographic parity, fairness through unawareness or group unaware, treatment equality. 

The concept of equalized odds and equal opportunity is that individuals who qualify for a desirable outcome should have an equal chance of being correctly assigned regardless of an individual’s belonging to a protected or unprotected group (e.g., female/male). The additional concepts “demographic parity” and “group unaware” are illustrated by the Google visualization research team with nice visualizations using an example “simulating loan decisions for different groups”. The focus of equal opportunity is on the outcome of the true positive rate of the group.

On the other hand, the focus of the demographic parity is on the positive rate only. Consider a loan approval process for two groups: group A and group B. For demographic parity, the overall number of approved loans should be equal in both group A and group B regardless of a person belonging to a protected group. Since the focus for demographic parity is on overall loan approval rate, the rate should be equal for both the groups. Some people in group A who would pay back the loan might be disadvantaged compared to the people in group B who might not pay back the loan.  However, the people in group A will not be at a disadvantage in the equal opportunity concept, since this concept focuses on true positive rate. As an example of fairness through unawareness “an algorithm is fair as long as any protected attributes A are not explicitly used in the decision-making process”.

All of the fairness concepts or definitions either fall under individual fairness, subgroup fairness or group fairness. For example, demographic parity, equalized odds, and equal opportunity are the group fairness type; fairness through awareness falls under the individual type where the focus is not on the overall group.

A definition of bias can be in three categories: data, algorithmic, and user interaction feedback loop:
Data — behavioral bias, presentation bias, linking bias, and content production bias;
Algoritmic — historical bias, aggregation bias, temporal bias, and social bias falls
User Interaction — popularity bias, ranking bias, evaluation bias, and emergent bias.

Bias is a large domain with much to explore and take into consideration. Bias and public policy will be further discussed in future blog posts.

This series of posts on Bias has been co-authored by Farhana Faruqe, doctoral student in the GWU Human-Technology Collaboration group.

 [1] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. CoRR, abs/1908.09635, 2019.
[2] Moritz Hardt, Eric Price, , and Nati Srebro. 2016. Equality of Opportunity in Supervised Learning. In Advances in Neural Information Processing Systems 29, D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (Eds.). Curran Associates, Inc., 3315–3323. 6374-equality-of-opportunity-in-supervised-learning.pdf
[3] Martin Wattenberg, Fernanda Viegas, and Moritz Hardt. Attacking discrimination with smarter machine learning. Accessed at, 2016

Discrimination and Bias

Our current public policy posts, focused on ethics and bias in current and emerging areas of AI, build on the work “A Survey on Bias and Fairness in Machine Learning” by Ninareh Mehrabi, et al. and resources provided by Barocas, et al. The guest co-author of this series of blog posts on AI and bias is Farhana Faruqe, doctoral student in the George Washington University Human-Technology Collaboration program. We look forward to your comments and suggestions.

Discrimination, unfairness, and bias are terms used frequently these days in the context of AI and data science applications that make decisions in the everyday lives of individuals and groups. Machine learning applications depend on datasets that are usually a reflection of our real world in which individuals have intentional and unintentional biases that may cause discrimination and unfair actions. Broadly, fairness is the absence of any prejudice or favoritism towards an individual or a group based on their intrinsic or acquired traits in the context of decision-making.

Today’s blog post focuses on discrimination, which Ninareh Mehrabi, et al. describe as follows:

Direct Discrimination: “Direct discrimination happens when protected attributes of individuals explicitly result in non-favorable outcomes toward them.”  Some traits like race, color, national origin, religion, sex, family status, disability, marital status, recipient of public assistance, and age are identified as sensitive attributes or protected attributes in the machine learning world.  It is not legal to discriminate against these sensitive attributes, which are listed by the FHA and Equal Credit Opportunity Act (ECOA).                

Indirect Discrimination: Even if sensitive or protected attributes are not used against an individual, still indirect discrimination can happen. For example, residential zip code is not categorized as a protected attribute, but from the zip code one may find out about race which is a protected attribute. So, “protected groups or individuals still can get treated unjustly as a result of implicit effects from their protected attributes.”

Systemic Discrimination. In the nursing profession, the custom is to expect a nurse to be a woman. So, excluding qualified male nurses for nursing position is an example of systematic discrimination. Systematic discrimination is defined as “policies, customs, or behaviors that are a part of the culture or structure of an organization that may perpetuate discrimination against certain subgroups of the population”.                                                                                                                              
Statistical Discrimination: In law enforcement, racial profiling is an example of statistical discrimination. In this case, minority drivers are pulled over more often than white drivers. The authors define “statistical discrimination is a phenomenon where decision-makers use average group statistics to judge an individual belonging to that group.”

Explainable Discrimination: In some cases, “discrimination can be explained using attributes” like working hours and education, which is legal and acceptable as well. In a widely used dataset in the fairness domain, males on average have a higher annual income than females because on average females work fewer hours per week than males do. Decisions made without considering working hours could lead to discrimination.                     

Unexplainable Discrimination: This type of discrimination is not legal as explainable discrimination because “the discrimination toward a group is unjustified”. Some researchers have introduced techniques during data preprocessing and training to remove unexplainable discrimination.   

To understand bias in techniques such as machine learning, we will discuss in our next blog post another important aspect: fairness.

Bias, Ethics, and Policy

We are planning a series of posts on Bias, starting with the background and context of bias in general and then focusing on specific instances of bias in current and emerging areas of AI. Ultimately, this information is intended to inform ideas on public policy. We look forward to your comments and suggestions for a robust discussion.

Extensive work “A Survey on Bias and Fairness in Machine Learning” by Ninareh Mehrabi et al. will be useful for the conversation. The guest co-author of the ACM SIGAI Public Policy blog posts on Bias will be Farhana Faruqe, doctoral student in the George Washington University Human-Technology Collaboration program.

A related announcement is about the new section on AI and Ethics in the Springer Nature Computer Science journal. “The AI & Ethics section focuses on how AI techniques, tools, and technologies are developing, including consideration of where these developments may lead in the future. It seeks to promote informed debate and discussion of the current and future developments in AI, and the ethical, moral, regulatory, and policy implications that arise from these developments.” As a Co-Editor of the new section, I welcome you to submit a manuscript and contact me with any questions and suggestions.