Join us on Thursday, April 8th at noon PST for a seminar about Urban Informatics with presentations by Dr. Arya Farahi and Dr. Kate Starbird.
Zoom link: https://umich.zoom.us/j/96580647748
“Quantifying and Mitigating Sources of Bias in a Decision-Support System”
Abstract: Applications of AI decision-support systems are increasingly shaping the fabric of our society. These systems can exhibit and exacerbate undesired biases that might hurt the under-represented population. Therefore, it is critical to evaluate these systems not only from a lens of predictive power and the rate of error but also from a lens of trustworthiness and fairness. In this talk, I will focus on two specific sources of bias in a decision-support system and propose mitigation strategies. In the first part, I will discuss biases originated from historical decisions and are reflected in data. I propose a metric of quantifying disparity in data and illustrate how we can alleviate these historical biases by applying simple modification to a decision-making system. In the second part, I will shed light on biases that are originated from predictive models. Predictive models are a central part of any decision-making system. The end-user act based on the information provided by these models. Biased or untrustworthy information mislead the end-user or incentivize the public to mistrust the system. I will present our mitigation method KiTE. KiTE is a hypothesis-testing framework with provable guarantees that enables practitioners to (i) test whether a model provides trustworthy information with respect to each sub-group of a population and (ii) estimate and correct for prediction bias at the individual and group levels.
Arya Farahi is a Data Science Fellow at the Michigan Institute of Data Science (MIDAS) at the University of Michigan. Their research contributes to the fields of astroinformatics and urban informatics; and is focused on understanding and mitigating the unexpected and not-well understood consequences of AI models, including algorithmic bias and uncertainty quantification, in real-world settings. Farahi’s Ph.D. training was in astroinfomatics with focus on developing computational and algorithmic solutions for learning and inference in uncertain settings. In the past four years, they simultaneously pursued urban informatics as a new research direction.
“Revealing the ‘Big Lie’: Methodological Innovation for Rapid Response to Online Disinformation”
Abstract: Abstract: In this talk, I’ll present preliminary research results from ongoing efforts to understand the spread of disinformation about the 2020 Election. First, I’ll describe the mission, structure, and everyday work practices of the Election Integrity Partnership — a multi-stakeholder collaboration that addressed mis- and disinformation about the 2020 U.S. election in (near) real-time through rapid response data science. Next, I’ll take you through some of our analyses to show how the “Big Lie” — the sustained effort to sow doubt in the results of the 2020 election — took shape on social media platforms throughout the latter half of 2020. I’ll highlight the participatory nature of this disinformation campaign and reveal some of the “super spreader” accounts that helped produce and sustain it. Finally, I’ll note how some of the social media platforms have evolved their strategies to address this kind of disinformation and wrap up by talking about what might come next, both in terms of platform policies and future collaborations for rapid response to disinformation.
Kate Starbird is Associate Professor of Human-Centered Design and Engineering at the University of Washington, as well as Data Science Fellow at the eScience Institute. The foundations for her research lies in the fields of computer supported cooperative work (CSCW) and crisis informatics — the study of how information-communication technologies are used during crisis events, including natural disasters (like earthquakes and hurricanes) and man-made disasters (such as shooting events and acts of terrorism). Primarily, Starbird’s work has examined how people use social media platforms during crises and other “mass disruption” events.
The Data Science Coast to Coast (DS C2C) seminar series is co-hosted by the eScience Institute, along with data science institutes at New York University, Rice University, Stanford University, University of California – Berkeley, University of Michigan and the Academic Data Science Alliance (ADSA).
In the first half of 2021, we will host five seminars, each featuring one faculty member and one postdoctoral fellow from two universities. Each speaker will give a 20-minute talk about ongoing projects and motivating issues, followed by 20 minutes of discussion with the audience. These seminars will be the launching point for follow-on research discussion meetings which will hopefully lead to fruitful collaborative research.