What makes a data scientist?
- 07 May, 2013 09:25
Ask a dozen CMOs or CIOs what tops their list of strategic priorities and odds are exceedingly good that "big data" ranks either first or second. One of the greatest challenges, they'll tell you, is finding the talent they need to analyse and wring business value from the ever-increasing volume of complex data flooding their enterprises. What they need, they say, are good data scientists -- and lots of them.
In one of the most frequently cited reports on the topic, the McKinsey Global Institute estimates there will be a shortfall of 190,000 data scientists just in the US job market alone by 2018.
But how exactly do you become one of these in-demand big data specialists? Is it a matter of training, certification or both? Is it simply the next logical career step for a traditional business intelligence expert? Is a computer science degree required?
As it turns out, there is no one right answer, at least not at the moment. Instead, it's largely a scramble out there on the big data field.
"Big data is like a kids' soccer game. Everyone is running to the ball, but no one knows exactly what to do with it. It has created a huge competition for people," says Greg Meyers, CIO at Biogen Idec in Weston, Mass.
"It's a very fluid area," agrees Michael Rappa, executive director of the Institute for Advanced Analytics at North Carolina State University. "Depending on what industry you're in or what company you talk to, it's a different reality when you talk about big data."
While a single definition might be elusive, academic, career and business experts agree that there are certain fundamental tasks all data scientists need to perform and certain skills are required to perform them well. The main pillars of the discipline are data clustering, data correlation, data classification and anomaly detection.
Or, as Rob Bird, a data scientist and CTO at Red Lambda, a provider of predictive security analytics, puts it, "You make data simpler, find relationships, find the weird stuff, and then make predictions."
Data Science vs. Business Intelligence: What's the Difference?
The terms "data science" and "business intelligence" seem to be used a lot in connection with big data, but they're really very different disciplines. Experts say data science is all about predicting the future, while BI involves producing static reports.
"Traditional BI engineers are effectively reporting information as is, even if they're reporting trends and standard deviations away from the norm," says Andrew Dempsey, director of DVD BI and analytics at Netflix. "They aren't really discovering new nuggets of information. The data is what it is."
But with data science, there's an element of mystery. For example, Netflix looks at historical data "to identify why someone is more or less likely to churn because of their behaviour," Dempsey explains. "There's more uncertainty there because on an aggregate level, a lot of people may have similar viewing habits, but on an individual level, everyone is different."
Another key difference between the two disciplines has to do with the data itself.
First, there's the sheer volume of data. "With so much data, you need to assimilate it to look at the exceptions, rather than the reports," says Biogen Idec CIO Greg Meyers. The pharmaceutical manufacturer, he says, continually reviews data from signals throughout the manufacturing process to detect when events are out of tolerance levels. When an anomaly is detected, a different operating procedure is triggered. "It's all about trying to make sure the process of how we manufacture is as controlled as possible," Meyers says. "We've matured our analytics process by looking at data across batches so we look at trends to reduce the variability of certain things."
Another challenge is dealing with the variability of big data.
Josh Williams, a data scientist at Kontagent, notes that "in classic BI systems, you usually have highly structured data -- things like customer profiles. You come up with an analysis by correlating that data and running regressions on it."
In today's big data environment, in contrast, "you have a mess of complex data and you have no idea how the features you may be looking at -- the input factors -- relate to the output," Williams says. The upshot is that data science is "much more exploratory. It's easier to shoot yourself in the foot. You have to be much more rigourous. It's much more difficult to do the analysis, which is why there is so much more research around machine learning," he adds.
Stalking the elusive data scientist
It is widely conceded that it is virtually impossible to find all the necessary analytical skills resident in one human being. The non-hysterical in the bunch of executives from the IT Leadership Academy we spoke to have rationally concluded that rather than stalk a mythological life form -- a data scientist with all the skills required -- they should adopt an "ensemble" approach to the deficit in analytical skills.
Here's how Scott Friesen, director for marketing analytics and customer insights at Ulta Beauty, explains this idea: "You have to create a portfolio of talent within a team. For example, you might have someone who is a great statistician but doesn't know database query mechanisms. So someone else on the team does the SQL pulls for the statistician, who hands off to the best communicator. That is who communicates the message to the business."
Glenn Wegryn, director emeritus of operations research at Procter & Gamble, skinned the analytical talent deficit in a very innovative way. As part of a multipronged talent strategy, he scoured the enterprise for employees who had analytical training but weren't employed in analytical jobs. This was a rich source of affordable raw quantitative skill. And that should not be surprising. Just about every student participating in the 6th Annual EEIC Engineering Capstone Design Showcase at Ohio State University demonstrated the raw skills necessary to create value with data.
So forget about the data scientist bogeyman. If you are eager to create value with data, go out and repurpose an engineer. They will love you for it.
- Thornton A. May
Universities Step Up
The skills required to perform these tasks cut across traditional academic disciplines, including statistics, mathematics and computer science. This is why several US institutions, including New York University and NC State, offer specialised data scientist certification and degree programs.
"Data used to be something you collected. It had neat rows and columns," explains Rappa. "You ran experiments that were time-consuming, laborious and costly, and you didn't have a lot of data so you dealt with sample sizes."
Now, in contrast, " data comes streaming off of every touch point you have with employees, partners and customers," he says. "Big data is about taking all of that data together and using it to optimise business or inventory levels or to better target customers. That's the trick of the whole thing. You need people who are good at handling large volumes of data and have knowledge of math and statistics to analyse the data."
Recognising this as early as 2005, NC State created the Institute for Advanced Analytics, which pulls together faculty members from various disciplines and teaches data science "in a very integrated way," Rappa says. Students take technical courses in statistics, finance and business, and they learn communications and teamwork skills, which Rappa says "almost always trump the technical skills," as far as employers are concerned.
Teamwork skills are critical, he says, because "you can't wrap up all of the [data scientist] skills you need in a single person." (See " Stalking the Elusive Data Scientist.") Instead, data scientists typically work in teams. IBM, for example, mixes statisticians with MBAs in its Data Analytics Center of Excellence, which helps businesspeople determine what questions they need data to answer. The centre's goal is to generate revenue through a marriage of business savvy and analytics, says CIO Jeanette Horan. One project optimised sales coverage in the 170 countries in which IBM operates, yielding a 10 per cent performance improvement in territories where the models were applied.
Rather than completing a final thesis, students work in teams to complete practicum projects with live data from major companies, including GE and GlaxoSmithKline. Seventy per cent of the program's students come from the workforce, many of them sponsored by their employers. Most students have at least two years of on-the-job experience, and their average age is 29.
At NYU, the newly launched, two-year master of data science degree is also multidisciplinary, intersecting mathematics, computer science and statistics. This is because to do data science well, "you need to have expertise in all three," says Roy Lowrance, managing director of the university's Center for Data Science.
Lowrance emphasises that data scientists also require what he calls "application knowledge." Without it, "you have no intuition about what to work on and test, especially in business," he explains.
What Lowrance refers to as application knowledge, some other experts describe as domain expertise. But whatever you call it, all agree that it's absolutely essential for data scientists in the business world.
Because data scientists are charged ultimately with showing business value, knowing a particular business is critical "because there's a lot of nuance in each domain", says Josh Williams, a data scientist at Kontagent, a company that finds and identifies customer behavioural insights from social, mobile and Web data in real time.
"A data scientist is someone who is familiar with statistics and classical mathematical analysis, and they need a strong background in programming and computer science or at least the ability to get things done in a programming language," Williams says. "But they also need domain expertise around how to apply different automated analysis algorithms to a given domain."
However, he adds, "data science skills are not necessarily industry-transferrable" because the volume and complexity of data varies from industry to industry. "We're dealing with orders-of-magnitude greater volumes, but the really important part is that the data is much more rich and complex," Williams says.