User interaction patterns can indicate conscious and unconscious biases, for eg. a job recruiter viewing only male candidate profiles on LinkedIn. The group developed models to detect and measure bias through many such interaction patterns. Now, as a designer, how can we take those bias metrics and make them usable in a visual analytics system? Specifically,

How do we design an interface that optimally mitigates the captured bias?

We derived a design space that has 8 dimensions that can be manipulated to impact a user’s cognitive and analytic processes. I designed an example system called fetch.data to illustrate potential bias mitigation strategies. It is a hypothetical dashboard for job recruiters to screen candidate resumes for hiring.
‍
Intervention strategies for bias mitigation:
1. Where and when do we present the bias metrics? [B.2: on-demand view of bias metrics]
2. How do we promote user awareness in order to reflect on their interactions? [A.1: increase the size of the data points that have not been interacted with]
3. How can we change the direction of their interaction upon detecting bias? [B.1: if recruiter interacts with only one gender, disable the gender filter]
4. What should be the degree of guidance to mitigate bias? [D: recommend candidates that have been excluded]
5. How do we collect user feedback to inform the model? [E: pop-up allowing recruiter to provide feedback while dismissing a notification]

A belief is an idea or feeling about the truth of a statement. In the context of data, beliefs can take the form of a hypothesis about a trend in the data. Eliciting a person’s beliefs about data is an increasingly common practice in contexts such as data journalism, where recent studies have shown that eliciting beliefs about the data leads to greater engagement, reflection, and memorability about the data. Belief elicitation may also be useful in a number of other contexts. For instance, in data analysis, a system that is aware of a person’s prior beliefs about the data may present things in a different order or format to mitigate the likelihood that a person misinterprets the data (e.g., due to confirmation bias around their existing beliefs). 

I lead this project, where I collaborated with Prof. Emily Wall (Emory University) and Prof. Yea-seul Kim (UW-Madison) to derive a design framework for belief elicitation to assist visualization creators like designers and data journalists in their design process.Our paper was accepted in EuroVis 2022, and I presented it in Rome in July.
‍Here's a simple example constructed by following this framework.