Since launching the UBC MDS program in 2016, we’ve received a lot of questions on why we designed MDS the way we did. The post will address the following design decisions:

  • Statistics and CS as the home departments.
  • Goal of the program: responsible use of DS.
  • Length of the program: 10 months.
  • Length of the courses: 4 weeks.
  • Creating all new courses from scratch.
  • The program prerequisites.
  • Dividing the instructor role into two pieces: lecture and lab.
  • Setting a single deadline for all weekly assignments.
  • A few words on the curriculum.

Stat-CS partnership

The UBC MDS program is an equal partnership between the Department of Statistics and the Department of Computer Science. I have noticed that some other DS programs lack the statistics component, and I think that is a big loss. To us, data science is not just computer programming and machine learning (and certainly not just deep learning!). Rather, it is a broad field about how to ask and answer questions using data.

As an aside, the collaboration between departments has been extremely productive and pleasant. When I heard that multiple departments would be involved, I was expecting some of the politics that universities are so famous for. While I found some of that in other places, the CS-Stat collaboration has worked great for MDS. It probably helped that both departments are part of the Faculty of Science, with a very supportive Dean.

Program goals

MDS is a professional Master’s program and we do not aim to produce researchers. Rather, we want to produce graduates capable of creative, responsible, and reproducible use of data science tools. Establishing this mission has been critical to fitting the program into only 10 months (see below), and has helped us sharpen our focus while creating our courses.

There is often the temptation as an instructor – and demand from some students – to go ever deeper into a topic. Our approach has been to reserve lecture time for what is essential to our mission, and to sprinkle our lab assignments with optional questions that go above and beyond our core curriculum. The students seem to like this, and it’s an effective strategy for dealing with such a diverse cohort of students. For example, a student with more databases background might complete the tough optional questions in the databases course; a student with a generally strong background may do more optional questions at the start of the year, when much of the material is review for them.

10-month program

For many MDS students, the program’s compact length is a major draw. For some, especially those leaving well-paying jobs to train in data science, a longer program simply would not be financially viable. (Of course, we also get many students who would prefer a longer program – it’s hard to please everyone!) Fitting everything we wanted to teach into such a short window was a big challenge, but also an opportunity to reflect on what content we felt was truly essential.

Some have also asked why we chose 10 months instead of 12. That is not for the students, but for us! The MDS team needs time to recharge every year, which is usually spent on a combination of vacation and upgrading the curriculum before the next cohort arrives.

4-week courses

Our program structure is typical in the sense of students taking 4 courses at a time, but atypical in that courses only last for 4 weeks. The benefit of short courses is that we can have more fine-grained control of how much we teach in a given area. For example, we have one course in algorithms and data structures, two courses in visualization, and three courses in supervised learning (depending on how you count). While we would love to have three courses (equal to a standard length UBC course) in every area, there simply isn’t time in 10 months. The 4-week structure also makes it easier to bring in subject matter experts to teach a short MDS course. The difficulty with this system is a potential lack of continuity between courses and instructors (more on this below), but overall I’m very happy with this structure.

For a glimpse of how the schedule looks on the students’ end, see the student-facing calendar.

24 new courses

The MDS program is structured as 6 “blocks”, each containing 4 of these short courses, for a total of 24 courses. All of these courses were designed and created from scratch for the MDS program. As you can imagine, 24 formal UBC course proposals amounted to a lot of paperwork – the whole package was over 200 pages – but it was worth it! We (and others!) feel that creating new courses is essential: because data science is such a rapidly evolving field, existing courses often weren’t relevant or sufficiently up-to-date. In designing these courses, we took a problem-first approach: what do these students need to be able to do in the workplace? I recall some treasured moments in which prospective employers visited campus and listed the skills they were looking for, only to discover they were essentially listing the contents of the MDS curriculum.

How did we organize this large-scale course creation process? Once the rough course titles were chosen, we set up a “first reading” for each course, in which the committee would review the overall course learning goals and a short title for each lecture and lab. Later, we would have a “second reading” to review lecture-by-lecture learning goals and rough drafts for all of the lab assignments. This process took place mainly over the summer of 2016, with a large group of people involved in various ways: Alex Bouchard-Côté, Jenny Bryan, Giuseppe Carenini, Paul Gustafson, Ed Knorr, Ed Kroc, Gail Murphy, Raymond Ng, Matías Salibián-Barrera, Tiffany Timbers, and myself.

By the way, all of this development was done with GitHub, and we would tag a release after each “reading”, which makes it easy (and fun!) to go back and view the development history of our courses. For more on how we use GitHub in MDS, see here.

Instructor roles

MDS was set up such that each course has two instructors: a lab instructor and a lecture instructor. All the labs are taught by members of the core MDS teaching team (i.e., people for whom MDS is roughly our full-time job). Having a core MDS instructor, who can see the big picture of the entire MDS program, in each course provides much-needed continuity across so many short courses. As for the lectures, some are taught by core MDS instructors as well. For the remaining lectures, the original plan was to bring in experts in each field from the pool of faculty members in the home departments (CS and Statistics), or in some cases from industry.

This seemed like a good idea given that data science is so broad – it’s hard to be an expert in everything! In fact, the saying goes that an MDS instructor can teach 1/3 of the courses any day, another 1/3 with some preparation, and the remaining 1/3 we should take as a student! I often like to say that the only people who know everything in the MDS curriculum are the MDS alumni. We teach a lot of things!

Despite all the arguments above, this teaching model has been problematic at times. With a distinct lecture instructor and lab instructor sharing each course, the potential for miscommunication is high, often resulting in frustrations for the students. Furthermore, we have a specific teaching style in MDS – very applied and hands-on – and it’s not reasonable for non-core lecture instructors to immerse themselves in our methodologies for just 4 weeks’ worth of teaching. Our new approach is to build the core team to include a wide range of expertise, so that we can staff an increasing number of courses entirely with core MDS instructors rather than bringing people into MDS for a few weeks at a time. While doing so, we also continue to benefit from the ideas and course materials created by the original set of instructors, and we continue to pick their brains for new developments in their respective fields.

In case you’re wondering who we are: when MDS launched, the core teaching team consisted of Tiffany Timbers (my fellow Option Co-Director) and myself. Since then, we have been very fortunate to add Vincenzo Coia, Giulio Valentino Dalla Riva (now at the University of Canterbury), Varada Kolhatkar, and Rodolfo Lourenzutti to the core team.

Program prerequisites

The prerequisite courses for entering the MDS program are listed online, but here is the summary:

  • One university course in programming.
  • One university course in probability and/or statistics.
  • One university course in calculus or linear algebra (completion of both strongly recommended).

On the one hand, we wanted to keep the prerequsites light to encourage applicants from a variety of academic backgrounds. This has worked out well; for example, see here and here for some demographic information about our 2nd and 3rd cohorts. It sounds like the students really benefit from being part of a diverse cohort.

On the other hand, if the prerequisites are too light, we might end up admitting students who aren’t ready to tackle the program. As one strategy for helping prospective students calibrate their background knowledge, we created a self-test (answers here) for students to self-assess their readiness. We also posted a list of MOOCs that might help them prepare.

One tricky issue we’ve encountered is “overqualified” students. Recently, our strategy has been to admit these students but add an extra paragraph in their offer letters saying something like, “You may already know a lot of this stuff… are you sure you want to enrol in MDS?” As a rule of thumb, I have found students claiming proficiency in ~1/3 of the courses tend to thrive in the program, but if we get to ~1/2 of the courses I would probably advise them not to join.

A unified assignment deadline

In the first year of the program (2016-17), each lab assignment was due at the start of the next lab a week later. As a result, students had four lab deadlines per week, plus quizzes. We found that students struggled with the large number of deadlines and with always having four assignments hanging over them. Another problem was that the students always had something “due tomorrow” which took priority over scheduled instructional activities. For example, in a DSCI 571 lab on a Wednesday, students would often be working on the DSCI 532 lab due Thursday, instead of getting help from the DSCI 571 instructor and TAs who were available to help them with DSCI 571 content due the following Wednesday. (No judgement here – I may well have done the same thing in their shoes!)

As a result of these difficulties, we switched to a single, unified deadline. This year, all the weekly lab assignments are due on Saturday at 6pm. We noticed a marked improvement in how lab sessions are used. This system has another major advantage: the students have no assigned work from Saturday 6pm until Monday morning. We feel that free time is extremely important for student wellbeing and productivity, especially within 10 months of intense study.

While this all may seem like a trivial matter of scheduling, I think it’s actually very important and would encourage other programs to deliberately carve out free time for the students.

The content

To learn more about what we teach in MDS, check out this blog post on the CS / machine learning stream of our curriculum. We are planning two more blog posts, on the statistics stream and “data science” stream; I will add links here when they are posted. Beyond these three streams of MDS courses, the 2-month capstone project is also a critical part of our program. For more information of our capstone course, see our capstone homepage, our partners page and various posts on this blog. For further information, see also a list of our courses with short descriptions and publicly availably teaching materials for some of our courses.

Mike Gelbart is Option Co-Director of the MDS Vancouver program and an Instructor in the UBC Department of Computer Science.