Some teams/hiring managers love them, and some candidates really detest them. Having been on both sides of the hiring equation, I can definitely see both perspectives. I wanted to write down some thoughts after using data/software challenges to help with the hiring of three interns at work.
Data/software challenges are take-home interview assignments that are given a specific time frame to be completed.
In the case of the three interns that I helped with hiring, two were for data science roles, and one was for a software engineering role. In all three cases, we gave them a one week time frame to complete the assigned challenge.
For the software challenge, the candidates were asked to build the core of a piece of typical software that we would build in our department, according to a pre-defined spec, i.e. pretty much in-line with how we operate, just on a mini scale. For the data challenge, the candidates were given a list of four data sets, and told to come up with a research question and produce an analysis that were delivered in the form of a Jupyter notebook.
After the data and software challenges were turned in, we would review them, and invite the subset of them that "passed" the bar to deliver a presentation "on-site". (With COVID-19, "on-site" has dramatically changed, of course.) The candidate had to deliver a seminar presentation on their challenge to myself and other colleagues involved in hiring.
The challenges were designed in such a way that a competent and well-equipped individual could complete the task in about 2-3 afternoons of work. We put in some thought to designing the challenges such that they could not be trivially solved by looking up StackOverflow answers alone. (Use of SO was encouraged, though, as long as they documented which SO post they referenced for the non-trivial pieces of the project.)
I can speak more for the data challenge, as I designed it.
To start, my data challenge borrowed ideas from the legend of Van Halen’s Brown M&Ms contract clauses: they contained some pretty exact requirements that seemed trivial, but if ignored, I would fail the candidate and not bother with their candidacy going forward. Being detail-oriented is a pre-requisite for doing data science work, and if the candidate did not follow instructions precisely, even after I factored in language ambiguities and barriers as confounding factors, I knew I would have problems working with them. In the interest of being able to reuse and evolve the data challenges, I’m clearly not going to reveal what those requirements are here, but the principle stands, nonetheless!
Apart from that, I was looking for a candidate’s ability to write “good” code (refactored and documented at least), take care to make sure their notebooks are “readable”, and be able to “defend” in their prose their choice of problem and the individual choices made in their analyses. As a bonus, I was also on the lookout for “modern” tooling. Having one or two grammatical or spelling warts was fine; we all make these mistakes once in a while, but if the grammatical issues interfered substantially with readability, then that was cause for concern.
Primarily, this list:
Meticulousness is most evident from their ability to pass the “Van Halen details” test.
Creativity would show in their problem choice and choice of methods.
Logical rigor would show up in how they defended their choices.
Empathy shows up in the way they communicate, both written and oral.
I am confident that these are non-controversial qualities to desire in colleagues.
This next statement will possibly be controversial, but I hold it to be true: taken holistically, one can tell from a presented pattern of good/bad behaviours whether a candidate can hit the quality bar for the points listed above. This is controversial because the desire for objective fairness in hiring means subjectively perceived patterns can be unfair. However, if there is sufficiently diverse perspectives agreeing on the same set of observed patterns presented, this cannot be ignored, regardless of whether the patterns present as positive or negative.
Possibly, though I did set a high bar because I expect the interns I work with to deliver high quality work. I’m ready to teach them things they don’t know, but a baseline level of skill level and quality is necessary. In order to ensure that my expectations were not set unreasonably high, I calibrated them against colleagues’ opinions; in particular, I asked them specifically whether I’d be asking too much of the candidate at their level.
Subjectively, I would claim yes, but because I ended up working with these interns, I might clearly be biased.
If my subjective experience is informative for anything, once I saw the data challenge notebooks for the interns I interviewed, I immediately knew whether I could work with them or not. In particular, I could get a pretty strong idea of their ability to write code well, which I prize very highly, as it is a strong indicator of being able to impose sane structure on an otherwise free form problem. Also, their ability to communicate clearly in prose (which has become important in a remote world) can shine through a well-done Jupyter notebook.
In one case, the intern we hired was the only plausible candidate available; the other candidate didn’t do the challenge.
In another case, I had the chance to evaluate three data challenges. One candidate failed the Van Halen test. The other two passed the data challenge with flying colours, and after interacting with them in person, I submitted my preference for the one who was able to defend their modelling choices in the live presentation. It was with that candidate too that I was able to go deep into technical points.
In all cases, I made sure to take the time to provide feedback to the candidates that we interviewed on-site. I remember not receiving concrete feedback post on-sites and being left frustrated, so I wanted to make sure this frustration was not passed on to them. My feedback was given right after I read the data challenge, i.e. before they came onsite, as I didn't want it to fall off the radar.
Possibly, though not always.
I floated the idea with my home team, but I did receive a legitimate concern: issuing a data challenge would be biased against those with families to take care of. After all, doing a data challenge does take time. So unless "not having a family to take care for" is part of the job description, I would be wary of making a data challenge an absolute requirement.
In the data scientist roles at NIBR, I’ve noticed that for whatever reason, we’ve prized prior scientific track record highly. That means either publication record or prior work experience. A PhD usually gives a candidate both simultaneously, so most of our data scientists are trained scientists who have finished doctoral training. (A small but excellent cadre of them are MSc scholars but have a significant track record too, e.g. in publications.)
The data challenge helps level the playing field a bit for MSc scholars, in my opinion. Without a track record to demonstrate creativity and logical rigor, the data challenge gives an outlet for that to happen. As such, for a regular colleague, I see it as being fair to offer the option to either present prior research work or to tackle the data challenge, without compulsion to do one or the other. Both should reveal the qualities that we’re looking for, and candidates that don’t have a prior track record can use the data challenge to shine.
The most important is to know what you’re looking for from the data challenge. It’s also helpful, I think, to spell the qualities expected to be demonstrated inside the text of the data challenge, to help set expectations on the candidates’ side. Doing so is also a matter of being thoughtful and intentional with the hiring process. If you don’t know who you’re looking for, how are you going to find that person?
Keep in mind that there are other ways to fish out the qualities you're looking for. An online portfolio that has been professionally curated, whether on a personal website or on GitHub or LinkedIn, for example. Hiring teams can look at the composition and rougly "grok" whether the candidate is competent, has the ability to write good code, has the ability to communicate (e.g. via teaching), and more. A data challenge should only be one component of the overall hiring process.
For a data scientist, I sense that "basic good software skills" are increasingly going to be important. That said, a data challenge is not the only way to assess a candidate's regular coding skill. Live coding challenges, which are done under artificial, pressure-cooked situations, are definitely not a good indicator. If there is code associated with a scientific publication, that is another way of assessing their software skills. Open source software development is another way of assessing this aspect. Finally, original code snippets embedded in publicly-delivered talks and/or blog posts are also an alternative way of getting a glimpse at a candidate's programming skills, though I find that the signal-to-nosie ratio there might be a bit lower.
I would advise against data/software challenges being a screening tool pre-phone screen. A data/software challenge is too much work to ask for up-front, and in my opinion a bit cruel if used to figure out whether to talk with someone. Adhering to the Golden Rule, I’d prefer to deal with a human before having to deal with a system, not the other way around.
The most important thing is to view it as another opportunity to shine, and not another hurdle to cross. This subtle switch in attitude is what can make a difference in whether your best side shows through in the challenge or not. The skills on which you can shine might include:
Be careful to see if there are Van Halen tests inside the data challenge. If there are, you may be unnecessarily disqualifying yourself.
Apart from that, if you do not see the qualities that a team is looking for, then that’s actually a red flag - it means they haven’t given enough thought into who they’re looking for. This could be an indicator of whether the entire data science team is being thoughtfully built out or not.
I hope this article is informative! It’s been on my mind to write down these lessons learned, and finally I found a chunk of time to do so. In the right hands, and paired with the right processes, data challenges can be an extremely revealing way of identifying whether an individual is someone you can work with. It shouldn't be the only or primary tool relied on. When hiring, I we should be on the lookout for people with whom we holistically can work with and who also complement the local environment of the team. An ecosystem of hiring tools are needed.