Data Science First Steps
3 Ways to Break Into Data Science
Starting your journey into Data Science? Choose the right path to increase your odds
With the popularity and demand for data scientists, and the well-documented shortage of skilled labor, more people are interested in data science as a career. Over time, I’ve gotten an increasingly large number of questions regarding how to start out as a data scientist. Like many other roles, landing the first job is typically the hardest, as having some experience under your belt is mandatory for many employers. This can create a vicious catch 22: how do you land your first job if they all require prior experience?
In this post, I’ll try to give you some advice — based on my own experience moving into data science several years back, and my current experience managing a data science department, interviewing dozens of candidates and reviewing hundreds of applications every year.
What’s your background?
From my experience, people trying to start a career in data science can be split into three relatively distinct groups. It’s important to identify which of these you are most similar to, in order to figure out your best next steps.
- The STEM career change — These are people with an advanced academic degree in a technical/scientific field who may already have several years’ work experience in an adjacent field. As the hype around data science has grown, they’ve started considering the option of transitioning. They typically have a strong mathematics and research background and can follow the linear algebra and statistics behind machine learning models. They have experience reading academic papers and aren’t intimidated by the formulas. Their transferable skills can help them become good data scientists relatively quickly.
- The data science new grad — While it’s taken a few years, universities have started to address the industry demand and various faculties are now offering MSc programs in data science. Depending on the university, these might include the statistics, electrical engineering or industrial engineering departments. While these degrees can’t cover everything, they’re quickly becoming a gold standard for comprehensive data science training that a 3- or 6-month bootcamp can’t meet. A good program will also include a thesis (and publication/s), which gives the employer an opportunity to discuss your work in greater detail. Whenever interviewing new grads I deep dive into their thesis, making sure they understand alternative approaches, discuss why they made certain decisions and ascertain how they handle feedback. Due to the scope of a thesis, it’s usually a great way to evaluate how someone performs research and how well they really know their material, in a way that a Kaggle project they did a while back can’t achieve.
- The optimist — This is someone who hasn’t gone through formal data science training nor do they have an extensive statistics/math background. They may have several years’ experience in data analytics within a specific vertical (finance, healthcare, etc) and want to complement their current skills to gradually move into a data science role. In the past, several people turned to me for consultation about their possibility to be a data scientist in fintech or some other specific vertical. While business acumen and experience in the vertical is important, this is the wrong mental mindset. The commonality between data science roles in various verticals is significant — the tools and algorithms solve generic mathematical problems, not vertical-specific ones. It’s easier to teach a good data scientist about a new domain than it is to train a business analyst with domain knowledge how to program, teach them statistics and machine learning. If you want to be a data scientist — you want to be just that, not a fintech data scientist.
If you’ve read this far, you probably know that there are a lot of online courses teaching everything data science related. While those courses are fundamental and deliver a ton of content, the vast majority try to give the most practical information as fast as possible. This typically means you’re going to learn a lot of machine learning models but only get the 30K foot explanation of how the algorithm actually works. Many courses won’t complicate matters with complex math so they can remain accessible to as big an audience as possible. While it’s definitely possible to train models and ‘do data science’ without understanding the intricacies of the algorithm, your capabilities will be limited. With the trend of automated ML picking up, plugging in an algorithm and trying out a few standard options won’t require a data scientist in the near future. Like many other professions, data scientists too will need to keep an edge over automated systems to keep their jobs, which will typically mean a much deeper understanding of the algorithms.
Due to the very accessible nature of data science training and lack of standard required qualifications to practice data science, anyone who has undergone a 50 hour course can self-appoint themselves as a data scientist. As elsewhere, when a role is in high demand, supply will increase to meet the demand and an influx of new candidates will start moving in. To have a serious chance at making it in the field, a significant investment of time is required.
How to break into data science
There are different ways to gain the minimal experience and knowledge to get your first data science position. When hiring for a junior position, the interviewer is going to look for a few things:
- Do you understand the fundamentals and theory of machine learning?
- Do you have the necessary coding skills (usually Python or R)?
- Can you demonstrate both of these points (e.g. walk the walk, not just talk the talk)?
As a candidate, you need to remember that the company’s loss function is asymmetric — hiring a bad candidate can have a much worse outcome than turning down a good hire. This means that companies are going to be cautious about taking risks on someone lacking a track record. You need to help the hiring manager as much as possible to demonstrate that you’re a low-risk and high-potential hire. This also means that your chances may be relatively low and you need to be emotionally prepared for a lot of rejections before getting an offer.
There are 3 main ways to gain the theoretical knowledge and expertise necessary for your first role, and they can be combined in various methods:
- Masters Degree (with thesis) — As mentioned above, this is probably the gold standard for training today. While it can take 1–2 years, it is time well spent, especially if studying at a well known university. University pedigrees vary by location so it helps to understand what’s considered a good university in your vicinity.
- Bootcamp — these typically run 3–6 months for full time immersive programs and much longer if they’re part-time. It’s best to pay close attention to the financial incentive the program has in regards to your future career. In some bootcamps it’s very straightforward — you pay for the training. On the other hand, the best bootcamps will also offer Income Share Agreements. In this scenario, after the bootcamp is complete you pay them a percentage of your salary only if it is above a threshold. The agreement is usually in effect for 2–4 years and is capped (e.g. 1.5–2X the upfront tuition cost). In Israel, ITC and Y-Data operate in this fashion and put a bigger focus on assisting their students land their first role. Other bootcamps work by keeping you on their payroll for 2 years following the training period, during which you work on a project for their client companies (e.g. Experis Academy in Israel). The bootcamp pays your salary directly and pockets the difference between it and their outsourcing fee, while typically offering the employee an exit clause (which covers their training expenses).
Generally speaking, these bootcamps cover a wide range of topics and include theoretical machine learning knowledge, coding skills, statistics and (at least one) capstone project. As you can understand, different bootcamps have various levels of incentive to ensure your successful placement following their training. In some cases, it may be worthwhile to invest the time in a bootcamp, even if a fair chunk of the material is already known just to benefit from their assistance in landing the first position. - Online courses — the amount and quality of these courses has been transformational, enabling anyone around the world to learn from the top experts. The fact that such high quality content is now freely accessible to anyone has dramatically reduced the barrier to entry. At a very high level one can separate these courses into two types — intro level courses that try to cover a bit of everything in machine learning, and more advanced courses that dive deeper into specific areas. Several of the popular intro level courses can be completed in under 80 hours of dedicated effort. While this does require dedication (especially for something doing this on top of a full time job), it’s a relatively trivial time investment compared to many other high-paying professions (e.g. think of the time required to become a pilot, lawyer or doctor). I’ve seen a few applicants who put down Andrew Ng’s infamous Machine Learning course as their single training in the field. I agree that it’s a great course (it was the first one I took when transitioning to data science), but it was definitely not sufficient to qualify as a data scientist. You should be very wary of any course that claims to teach you the A-Z of ML. They might be a great intro into the field, but you should treat them as the first step in a long journey.
What do these trends mean for me?
The STEM career change — Of the three paths this is probably the fastest one, and if you invest enough time, your chances of success are pretty good. Additionally, the closer your background is to data science, the better. Depending on your background, you may already have most of the mathematical background and need to invest more heavily in your programming skills. As an employer, discussing someone’s thesis or dissertation can help show how well they grasp complex research subjects. Can they get into the weeds and back up to 30K feet quickly? Do they really understand why they made different decisions or used certain algorithms? What value might their research have? While strong research capabilities aren’t enough for a data scientist, checking these marks can help de-risk a new candidate, especially one with limited direct experience in the field. As someone who went through this path several years back (my MSc was in applied physics), I continue to see how my education gives me a different viewpoint in solving problems compared to colleagues with math, statistics, economics or biology backgrounds.
Someone going through this path also has the benefit of being able to pick up more advanced material quickly. Once you’ve gotten your feet wet, you’ll want to understand the algorithms to a great extent and develop an insight for the hyperparameters. This is a lot easier if you’re accustomed to advanced math.
Pro Tip — if you’re at all able to highlight data science / machine learning work you’ve done before you officially started as a data scientist, you might be able to get additional years of your experience recognized as relevant when negotiating compensation. While you don’t want to embellish your past work, it is useful to point out your programming experience, data analytics, advanced statistics, experimental design, algorithm development or other adjacent types of work.
The data science new grad — assuming you still have some time to complete your studies, look for any extra-curricular activities that can help you gain experience. Ideally, this would involve an internship within a data science team. One of my past employers would regularly bring in interns each summer and make offers at the end of the season to the most promising ones. This was a great win-win and a large portion of the company’s hires came through that program. If an internship isn’t possible, your university might have a capstone project you can invest in. At Riskified we’ve collaborated with a local university, giving one of their teams an open project to work on with our guidance as their capstone. If the students invest and do genuinely good work (i.e. not just to pass their course, but something that would qualify as good work in the company), we could be interested in hiring or at the very least writing a letter of recommendation for future employers.
Pro Tip — When working in data science (as in almost any career), you’ll need to be able to explain things to people outside your domain (side note — never make the mistake of thinking non-technical people aren’t as smart as you). During your interviews, you’re going to be asked quite a bit about your thesis. Find a smart friend with limited knowledge in machine learning to ask you about this. Can you explain to them what you did and how it was different from existing solutions? I’ve interviewed several new grads who could describe all the details of their research but were stumped by some high level, introduction questions (e.g. why is this research important?).
Finally, don’t forget that success requires lifelong learning and you’ve only completed one phase of your training so far. Continuing to learn on the job is just as important and may be more difficult as it isn’t as structured.
The optimists — There are a lot of people learning to become data scientists through online courses and bootcamps. Competition is stiff and you’re not going to get a job in the field after investing 80 hours. Employers are going to look at the duration of your classes/bootcamp and how familiar they are — nano-degrees on EdX or a 6-month bootcamp are going to be a lot more impressive than a single course on Udemy or Coursera.
In my opinion, the window of opportunity to transition into data science without extensive formal training (e.g. self-taught online courses) is shrinking. While it’s still doable, you need to realize that there are a lot of people with shallow knowledge of the field and landing your first job will require a lot more (as of September 2020 Andrew Ng’s course has had 3.5M enrolled students). If you want to go down this path, it will probably still take you several months (read: hundreds of hours) of course work and hands-on projects with a good dose of luck.
Pro Tip — if you can, consider bootcamps that have a proven track record of alumni starting data science positions (if their financial incentive depends on this, even better). While several months of full-time studying might be more than the investment you were considering it could make all the difference.
Due to the slow but steady autoML trend, it also means that you need to keep studying and increasing your expertise after you’ve landed your first role. You always need to stay a few years ahead of automation and a little bit of paranoia can be healthy for long-term job security.
Final thoughts
Compared to other high income, high demand professions, you don’t have to spend several years in medical school or log a thousand flight hours before you’re allowed to practice data science. While the demand for data scientists is high, most of that demand is for very skilled individuals who can demonstrate their value. You need to keep in mind that despite the lack of regulatory barriers, market forces still exist and companies won’t pay top dollar for someone with limited experience. More so, new data scientists require a lot of attention, training and support from more experienced data scientists. As the first few months are almost all investment by the company, it could take a year until a new data scientist’s contribution is back to zero. Paradoxically, this problem is exacerbated by the lack of experienced data scientists — they are really needed working on problems now and can only spend a certain amount of time training new people.
It’s not an easy path but it’s definitely rewarding. The world needs more great data scientists, so get to it!
— — — — — — — — — — — — — — — —
Thanks for reading! If you’d like to read more, you can follow us on our Riskified Technology publication
