Het begrijpen van vooroordelen in de representatie van mensen in algoritmische beslissingsnemers

This thesis explores the challenge of achieving representational adequacy in algorithmic decision-making (ADM) systems, particularly in contexts involving social outcomes. In machine learning (ML) using tabular data the attributes selected to represent individuals are operationalizations of individuals' characteristics and life circumstances. Representations are not replications and there is always a discrepancy between the representation and that which is represented. This discrepancy is referred to as bias (herein) if it contributes to ADM systems systematically and unfairly denying opportunity or resources or assigning an unfavorable outcome to individuals or groups. The concepts of fairness and bias are normative concepts about which differing stakeholders hold differing beliefs--this is as opposed to having a single definition that can be formalized for computational use. As such, this thesis focuses on the impact of including multiple stakeholders in the design and development of computational representations for ADM systems, with a particular focus on domain experts and affected communities as stakeholders. The two central research questions of this thesis are: (1) How does the incorporation of socio-technical analysis and justice-oriented participatory methods into computational representation design address the bias incorporated and identified in those representations? (2) What is required for CS researchers to incorporate these methods into their work in the face of inter- and trans-disciplinary differences? To address these questions, the thesis engages with the field of Fair Machine Learning (fair ML) and the interdisciplinary fields of Critical AI Studies and Science Technology Studies. To answer these questions, the research employs a mixed-methods approach. Qualitative data is collected through interviews, autoethnography, and co-design workshops and analyzed using thematic analysis. In addition, a computational method is proposed to address population misrepresentation in training data when using decision trees. The dissertation is structured as follows: Chapters 2 and 3 present case studies on ADM in Public Employment Services, illustrating the impact of stakeholder needs on system design. Chapter 4 introduces a novel ML model that incorporates expert knowledge to mitigate population-level representational issues. Chapter 5 addresses the challenges and considerations for researchers engaging with marginalized communities during the design of ADM systems. Finally, Chapter 6 analyzes the challenges of integrating interdisciplinary approaches in the field of Fair ML through the lens of the theory of doability. The three contributions of this work are (1) case studies on stakeholder-driven design in ADM systems, (2) a method for addressing distribution shifts in decision trees by incorporating expert knowledge, (3) an analysis of the inter- and intra-disciplinary challenges in this approach and guidelines directed towards computer science researchers.