Read on Twitter
Tomorrow, the UMD physics department is hosting a very interesting colloquium talk on the use of machine-learning for graduate admissions. I'd like to take a second and explain why this talk is concerning. Thread (1/13)
https://umdphysics.umd.edu/events/physicscolloquia.html
https://umdphysics.umd.edu/events/physicscolloquia.html
First, a quick summary: GRADE is a statistical ML system that uses historical admissions data to predict how likely the committee is to admit each new applicant. It's supposed to save time by helping reviewers focus on parts of each applicant’s file that matter the most. (2/13)
The first red flag is "historical admissions data". Diversity in computer science has always been less than stellar, and the use of historical admissions data to train the model implies a continuation of the racial and gender inequity from the previous years. (3/13)
I wasn't able to find the graduate student demographics for UT Austin CS, but their undergraduate demographics definitely paint a certain picture. And since graduate diversity is almost always worse than undergraduate diversity... (4/13)
If you want to learn more about diversity in computing, check out these links. In short, it's bad. You'd think a field that puts stats on a pedestal would have more documented statistics on its own demographics, but 
(5/13)
https://codeorg.medium.com/is-diversity-in-computing-jobs-improving-32f30068b7de https://www.wired.com/story/computer-science-graduates-diversity/
(5/13)https://codeorg.medium.com/is-diversity-in-computing-jobs-improving-32f30068b7de https://www.wired.com/story/computer-science-graduates-diversity/
Returning to the model: GRADE relies on a "small subset of highly predictive features" that prefer "applicants with high GPAs and test scores, backgrounds from reputable institutions, and recommendation letters that support the applicant’s potential as a researcher." (6/13)
In a vacuum, it makes sense to assume that high academic stats would translate to high performance in grad school. Unfortunately, we have never had the luxury of living in vacuum, and to pretend that we do is inherently racist. (7/13)
It has been shown repeatedly that GPA, test scores, and other similar stats are highly dependent on race, class, gender, and other factors. Many of these features are fraught with bias, and lack any strong correlation to graduate school GPA or publication rate! (8/13)
The applicants GRADE is effectively searching for are white men from "prestigious" PWI institutions like the Ivy Leagues. They're just twisting the science to hide it. This quote, I think, really sums up their ulterior motive. (9/13)
"Another interesting finding is that the applicant’s gender, ethnicity, and national origin receive zero weight when provided as features to the model. This result indicates that UTCS admissions decisions are based on academic merit." 
(10/13)
(10/13)
Did they think that by throwing that sentence in there with absolutely no statistics to back it up, that they could absolve themselves of racism and sexism? Clearly, UT Austin thinks so, and by giving this talk a platform, so does UMD. (11/13)
Remember that this paper was published in a peer-reviewed magazine. There are so many red flags strewn throughout the paper, and instead of being rejected, this paper made all the way to a colloquium talk in another field. (12/13)
Decades of education research cannot be replaced by an algorithm trained on biased data. This is bad science, but more than that, this harms BIPOC, gender minorities, and international scientists. How can you claim to care about DEI when you give these papers a platform? (13/13)
Also, the entire graduate school admission process is built to disenfranchise folks! See bonus meme:
