The School of Data Science and Society and the Department of Political Science in the College of Arts and Sciences are hosting a seminar featuring Michael Colaresi, associate vice provost for data science at the University of Pittsburgh.

Why Models (and Not Just Data) Can Be Biased and What that Means for Responsible Data Science and AI
Everything from package delivery to missile targeting and from learning foreign languages to Nobel Prizes has been touched and in many cases transformed by the availability of streaming data, flexible algorithms and pervasive connectivity. Yet, simultaneously, there has been growing recognition that digital tools can and often do have biases that distort information in ways that harm individuals and groups. The goal of this lecture is to explore in more detail the sources of biases in data-driven systems in order to improve downstream applications and decisions. There are many high profile examples in academic work, policy white papers, think tank position pieces and educational materials where bias in digital systems is attributed simply and solely to problems within the “input data”, “training data” or just the “data”. I argue that while misalignment of structured inputs with the intentions of an application is indeed possible (and even probable), models and algorithms themselves can and do also contribute to biased inferences and downstream harm. The presentation also offers a general framework for clarifying how evidence and assumptions jointly produce inferences (good, bad and ugly) and how an augmented Box’s Loop (Blei 2014) spotlights the pivotal role of data scientists in responsibly aligning machine performance with human principles. He will conclude with practical examples of their work improving this alignment in practical settings from government, industry and academia.

About Michael Colaresi
Michael Colaresi is associate vice provost for data science and leads the Responsible Data Science Initiative at the University of Pittsburgh. He is the William S. Dietrich II Professor of Political Science, assistant director of the Center for Research Computing, and co-founding director of the interdisciplinary computational social science major. His work leverages the accelerating availability of computational tools, including machine learning and Bayesian approaches, along with unstructured information, such as from digitized text, to build and improve models of national security problem-solving, international and intrastate violence and changes in human rights over time. He also develops computational and visual tools that enable domain specialists to work alongside computer scientists to improve specific applications. He was previously the research and academic director of the Institute for Cyber Law, Policy, and Security, co-editor of the journal International Interactions and was co-recipient of the Best Visualization Award from the Journal of Peace Research in 2017 and the Gosnell Prize for Excellence in Political Methodology from the Methodology section of the American Political Science Association in 2006. Colaresi’s research has been funded by four NSF grants and he is leading a large-scale effort to develop scenario-based training in responsible data science that is funded by the Richard King Mellon Foundation. He is an external researcher at the Peace Research Institute-Oslo and an external expert for the European Commission.