Within our community, research is dominated by quantitative study of observational data. Such an approach has provided important insights into algorithmic bias, but also cannot easily be used to study feedback loops that help produce this bias. This workshop will discuss new approaches to quantitative evaluation of observational data, as well as new quantitative experimental approaches, but also will emphasize the benefits of qualitative and simulation-based approaches. Qualitative study can help to identify, e.g., how algorithms are (or are not) put to work across difficult-to-quantify social contexts2. Similarly, while identifying problems in real systems through empirical measurements is a great first step towards accountability, we can not possibly investigate each system one-by-one, as they change over time, etc. using empirical measurements. To study these kinds of behaviors, it is instead useful to approach such a problem from a simulation perspective. Using simulation, we can assess how different potential changes to the system impact the evolution of platforms and their users over time.

In addition to the various methodological perspectives that must be brought to bear on this problem, this workshop also emphasizes the need for various theoretical perspectives. In many cases, applicable theory from the social sciences may not be known to, e.g., computer scientists, and conceptual tools from social physics used to describe complex systems may be unknown to both social scientists and computer scientists. To this end, this workshop seeks to bring to bear theoretical perspectives from across these disciplines that use a complex systems perspective to help explain or understand algorithmic bias.

Recent research has argued that the problem of algorithmic bias lies largely in structures of bias and discrimination that lay beyond the boundaries of model parameters. Further, developing a “fair” algorithm may be difficult, particularly because definitions of fairness are themselves varying and socially constructed. At the same time, however, at the scale of today’s social media, it is obvious that some automation is required if we wish to, e.g. provide friend recommendations on Facebook. How, then, can we move forward in a world where algorithms seem destined to be both necessary and that may never fully be "de-biased”?