Biases as Values: Evaluating Algorithms in Context

Díaz, Mark

doi:https://doi.org/10.21985/n2-k9ba-mb78

Work

Biases as Values: Evaluating Algorithms in Context

Public

Download PDF

Download All Files (.zip)

This dissertation asks how researchers can create more equitable algorithmic systems. Ultimately, this thesis explores methods and implications of representing subjects of analysis in the design and evaluation of algorithmic systems. I also unpack how algorithmic tools measure and quantify human behavior, giving heed to the potential impacts of these systems on underrepresented communities. Building off of current work in HCI and algorithmic fairness, my research raises questions about how we can evaluate algorithms to understand the contexts and communities they serve best. The data sets and survey data used in this research are available at https://dataverse.harvard.edu/dataverse/algorithm-age-bias/. Because algorithmic tools are often created using data processed in similar ways (e.g., using one of a few common, publicly available data sets, or generating training annotations exclusively through crowd work platforms), these tools can fail to capture data and patterns that reflect underrepresented groups. The result can be unintended algorithmic bias. I use mixed methods-- both quantitative and qualitative-- to explore the broader social contexts in which algorithmic tools are applied and test methods of mitigating unintended algorithmic bias---one of which directly involves subjects of analysis in the creation and evaluation of an algorithmic tool. In my first study I systematically analyze the outputs of popular sentiment models for age bias. I find a tendency for text to be rated more negatively when it references older age, then I develop an approach to removing bias rooted in training data. Taking older adults as a stakeholder group to prioritize in the face of age bias, I next solicit data annotations from older adults to evaluate model performance against their expertise on age and aging. A new model built from this data replicates age bias of similar magnitude to bias in my initial analysis and raises questions about the influence of annotator social identity and beliefs on model performance. In the final study I complement quantitative approaches to model assessment and turn to qualitative methods to evaluate model objective functions using direct input from algorithm stakeholders. Ultimately, I argue that the development of ethical algorithmic tools must involve input from the very individuals who will be analyzed and impacted by system deployment.

Creator