Skip to main content

Short version: Philip Guo and I wrote a paper about how a team of data scientists built their own software tools for making, maintaining, and iterating on an educational program they designed to help adults from historically marginalized populations to start their data science careers.

The real story of this research project starts when I was sitting in the audience at Shannon Ellis’ Design at Large talk. She was explaining Chromebook Data Science (CBDS), a small miracle of a program that is offering paid educational opportunities to adults in East Baltimore for training in data science. East Baltimore is a historically marginalized neighborhood where the effects of redlining can still be seen today. The goal of CBDS is to open up educational and career pathways in a community where such opportunities, especially in technical fields, have been historically scarce. Currently they find interested adult students in East Baltimore by working with local organizations like HEBCAC, an organization that helps East Baltimore residents to further their education or find jobs. Each student gets a tutor, they are given a Chromebook, and over the course of twelve weeks they learn how to use their Chromebook to do data science and cloud computing based work. After the program the CBDS team help place their students into local data science jobs.

I got inspired during Shannon’s talk about this new program when I realized that there were several stories with research implications that were simultaneously unfolding as part of the CBDS project. The central and most important story of course is the impact this program is having on the community it is serving, and the effect it will have on the student’s careers. The team building CBDS is in the process of iterating on their program while looking at the longitudinal outcomes of their students to evaluate the extent to which CBDS is effective. I am optimistic that they will have initial results to report on soon.

However, there is another story here about how the CBDS team was able to create their program is such a short amount of time (three months), how they are able to keep the course up to date despite how quickly data science technologies are changing, and how they were able to quickly adapt course materials based on feedback from their students. The CBDS team was made up of mostly data scientists who engage in end user programming, a genre of programming focused on using programming to achieve a goal, rather than to build a software system. As data scientists, the CBDS team had years of combined experience using programming to answer questions about biological and health data. To build CBDS the team repurposed the tools of data science (and the spirit of end user programming) to affect social change and to explore new methods of producing an educational experience.

To investigate how this happened we interviewed everyone on the CBDS team, asking them about the work they did for the project, and how they collaborated with each other. It turns out that they already had many of the tools and skills they needed to build this program from their experiences and expertise that they had developed doing open and reproducible science. Almost all of the course materials were written in R Markdown - the most popular computational notebook for R programmers. R Markdown is a plain text format, therefore the team could use collaboration and versioning tools like Git and GitHub. Several members of the CBDS team routinely write academic papers, blogs, and even interactive dashboards using R Markdown, so it was natural for them to extend this format to creating courses.

To transform these R Markdown files into the artifacts the team needed for teaching a course the CBDS team partnered with Leanpub, an eBook publisher, to launch a new online course platform. This platform can ingest modified Markdown files to create an online course structure with individual lessons and assessments. Data science is a rapidly changing field, so keeping data science courses up to date is usually challenging. Online course providers like Coursera only allow instructors to change the content of their course through a web interface, so making lots of changes to course has been historically difficult. This difficulty is especially frustrating in data science courses which can depend heavily on a few software libraries. If the API for one of those libraries changes, then updating the entire course can be extremely tedious. However, since all of the course materials for CBDS are based in plain text documents like R Markdown, updating materials to incorporate changing technologies is significantly easier and closer to the workflows the CBDS team uses in the rest of their work.

Videos and video lectures are a central component of many online courses, and the CBDS team confronted many of the same challenges with video content as they did with the rest of the materials in their course. Whenever the API for a software library would change, they would need to re-shoot and re-edit the video, a process that consumed valuable time and resources. In order to keep their videos up-to-date and decrease the amount of time required to produce each video they developed Ari, a software library for R that allowed them to write R Markdown documents that could be compiled into videos. To create a video using Ari, a team member would create a series of lecture slides using R Markdown. They would then write a script that would be read over each slide using text-to-speech technology that Ari could access. Ari would then stitch together the narration audio file and the series of slide images into a video. Since the source material for all of the course videos were plain text files, changes to videos could be tracked and videos could be created collaboratively using Git, GitHub, and the rest of the CBDS team’s typical workflow.

To manage and test all of these materials and digital artifacts that were being created for the course, the CBDS team developed Didactr, another R library that tested and packaged all of their materials and videos. The CBDS team was accustomed to the data processing pipelines common in genomics and bioinformatics research, and Didactr served as a comparable piece of pipeline-like software for proofing and publishing their online courses.

The sum of this entire process suggests the emergence of new technologies for maintaining educational materials which I have been thinking of as “EdDevOps” or DevOps for education. Traditional DevOps (which is still a relatively new field) combines software development practices with the operational and deployment requirements that are necessary for running that software. EdDevOps is similarly concerned with development, deployment, and maintenance, however applied to educational materials and specifically materials that address rapidly changing fields like data science. In the future instead of authors and instructors needing to painstakingly revise course materials, this process could be significantly aided by software that is empowered to update materials automatically.

End user programming means using programming to achieve a specific goal, and in the case of the CBDS team this goal was to affect social change in their local community through creating new educational programs and career opportunities. CBDS is of course not the first effort towards social change driven by end user programming. Data for Black Lives is one example of how data sharing and analysis can be used to empower social change. The case of CBDS is remarkable however because of how many tools and paradigms that were already available to the CBDS team could be repurposed to meet their new socially focused goal. This case study opens up questions about who is in the best position to create opportunities like CBDS, and how end user programming can be an empowering force to this end.