Towards Engineering AI Software for Fairness

Towards Engineering AI Software for Fairness: A framework to help design fair, accountable and transparent algorithmic decision-making systems

Lazo, Claudio (TU Delft Electrical Engineering, Mathematics and Computer Science)

Houben, G.J.P.M. (graduation committee)
Lofi, C. (mentor)
Venkatesha Prasad, R.R. (graduation committee)

Delft University of Technology

Computer Science | Web Information Systems

2020-08-25

Algorithmic decision-making (ADM) is becoming increasingly prevalent in society, due to the rapid technological developments in Artificial Intelligence. ADM make substantially impactful decisions about people: diagnosing whether we have a disease, what news and which ads we get to see, whether we
are eligible for a job, benefits, a college or a loan, they show us personalized media and news, and steer the car that drives us home. However, ADM brings about ethical, legal and social issues by inheriting and perpetuating human biases, learning to discriminate—even learning gender or racial stereotypes, and lacking transparency and accountability. This unexpected and biased behaviour arises because these software systems are usually built without the specification of fairness requirements (i.e. what fair behaviour is expected of the system). We envision a Software Engineering for Values (SEfV) method that solves this problem.
This study addresses that specification problem, aiming to help practitioners design ADM software for fairness. Using literature in social sciences—specifically organizational justice—the human value of fairness has been conceptualized in regard to ADM. This resulted in a fairness tree with four dimensions (procedural, distributive, informational and interpersonal fairness), which is further specified into 31 fairness norms. Subsequently, the fairness tree is related to current measures of fairness and techniques. Finally, we put forward the Software Engineering for Values (SEfV) framework, based on the principles of Software Engineering and Design for Values, and show how it can be applied to design ADM for fairness.
Experiments were conducted where participants (N = 12) performed a design task (M = 3, 75 requirements specified) and an audit task for a hypothetical loan decision system—using a prototype of the SEfV framework. Participants found the prototype useful for both design as auditing, especially
as a tool for reflecting on fairness considerations. This suggests that a high fidelity version would be useful for practitioners.

fairness
discrimination
bias
algorithmic decision-making
machine learning
software engineering
requirements engineering
Design for values
AI ethics

http://resolver.tudelft.nl/uuid:c7dc661e-c3f9-4986-bc54-c903aaddbc68

Student theses

master thesis

© 2020 Claudio Lazo