Detecting Bias in Intelligent Autonomous Systems (DeBIAS) (GD22)

About Us

    We are Team DeBIAS, which stands for Detecting Bias In Automated Systems. We are an undergraduate research group studying algorithmic bias in online hiring between consultants from historically black colleges and universities (HBCUs) and non-HBCUs.

  The goal of our research is to recreate a hiring algorithm, show that there is bias in the algorithm, and mitigate said bias. Currently, companies are unable or unwilling to develop their hiring model to be unbiased, causing issues for disadvantaged job seekers. We want to develop a low-cost method to detect and mitigate said bias so that companies can apply it to their own hiring models. 

What Will We Do?

    We are dividing our research into two different hypotheses: preliminary analysis and algorithm training. 

    For the first hypothesis, we are collecting resumes from Indeed, an online hiring website, along with their relative rankings from the first position on the starting page. We will analyze the information to see if there is a correlation between whether the applicant attended a HBCU and their respective ranking on Indeed. We hypothesize that on average, we will see resumes with an HBCU rank lower compared to resumes without an HBCU. These resumes will also assist us for the algorithm training in the second hypothesis.

 For the second hypothesis, we will use the resumes from Indeed to develop an algorithm which can emulate the performance of the Indeed algorithm that ranked the resumes in the first place. We hope to use this model and adjust it so that we get an even distribution of resumes with HCBUs and non-HCBUs on them across all rankings, therefore unbiasing it.

Why Help Us?

    Throughout our research process, we have discovered a variety of ways that “well intentioned” hiring algorithms have inadvertently become biased against disadvantaged groups. Learned biases in machine learning hiring systems may perpetuate and worsen existing social prejudices through our own confirmation bias. These biased systems often produce faulty outcomes, resulting in the system overlooking qualified applicants. This issue has recently come to the forefront of national news, with Amazon’s AI hiring system unintentionally perpetuating prejudice and injustice against women. 

    Our research will model the resume rankings of an online hiring website with the aim of determining how it ranks differing education levels and backgrounds. With this we can judge fairness, and if the website results produce any unfair outcomes. From there, we can evaluate different methods of debiasing the algorithm, including more data collection, a better balanced dataset, or manual model adjustments. These can inform future hiring algorithms of methods to avoid potential bias.

Gifts in support of the University of Maryland are accepted and managed by the University of Maryland College Park Foundation, Inc., an affiliated 501(c)(3) organization authorized by the Board of Regents.   Contributions to the University of Maryland are tax-deductible as allowed by law. Please see your tax advisor for details.