What is data science?¶

The combination of the words data and science do not give much hint as to what data science might mean. Most science has data, and most work in science is analyzing data.

Because the name data science can cover such a broad range of subjects, it has proved difficult to define.

We can make the definition very broad. For example, the UC Berkeley data science textbook has this definition

Data Science is about drawing useful conclusions from large and diverse data sets through exploration, prediction, and inference.

As you will see, we do not think this very broad definition captures what is really new in the data science movement. We prefer to define data science as a new culture of data analysis - like this:

Data science is an approach to data analysis with a foundation in code and algorithms.

Why do we choose this definition? And does it matter?

We start with the second question. If data science is important, then we need the right definition in order to chose what we do and teach.

Does data science matter?¶

Even if data science is hard to define, it has significant influence in planning for education, and science.

David Donoho is an eminent statistician at Stanford University. He recently wrote an article reflecting on data science, how it was defined, and what it could mean. He was worried that data science would be defined in a narrow way, in terms of big data and machine learning. He was worried because he thought that data science was important - that it should be:

… the really important intellectual event of the next fifty years.

[Don15]

For Donoho, data science is the rediscovery of the powerful methods of data exploration and analysis used and taught by statisticians such as John W. Tukey [Tuk62]. These methods will be revolutionary for the future of data analysis.

Meanwhile, educators, especially in the United States, have concluded that data science techniques are fundamental to the future of education. A 2018 report of the National Academies of Science Engineering and Medicine recommended that:

… academic institutions should encourage the development of a basic understanding of data science in all undergraduates

[NAoSM18]

The origins of data science¶

The phrase “data science” has been around for a long time [Pre13] but its meaning in current use is very recent. As it is currently used, the phrase “data science” comes from the job title “data scientist”, and the job title came from the tech industry. D.J. Patil, at LinkedIn, and Jeff Hammerbacher, at Facebook, noticed that they had a “new breed” of data analysts - these were scientists who could code. Thomas Davenport and Patil later described this phenomenon in a famous article “Data Scientist: The Sexiest Job of the 21st Century” [DP12]. One section of the article is “Who are these people”:

… what data scientists do is make discoveries while swimming in data … At ease in the digital realm, they are able to bring structure to large quantities of formless data and make analysis possible. … Data scientists’ most basic, universal skill is the ability to write code.

Most had come to code through their scientific work:

Some of the best and brightest data scientists are PhDs in esoteric fields like ecology and systems biology. George Roumeliotis, the head of a data science team at Intuit in Silicon Valley, holds a doctorate in astrophysics.

These people needed a new job title, because they were so much more effective than other data analysts. Because they could code, they were able to analyze a much wider range of data, and they could build programs to do difficult tasks such as the analysis of big data sets.

Soon after these articles, data scientist started to become a very valuable job title, with many companies competing to find people who could do this work.

A new culture¶

Data scientists in industry were not doing anything new, in an academic sense. They were applying skills that they had learned in academia. Central to these skills was the ability to build analyses in code. It turned out that doing this led naturally to many of the powerful practices discovered and described by statisticians like John W. Tukey [Tuk62] and Leo Breiman [Bre01].

What’s new in data science?¶

What are these practices, that make data scientists in so effective in industry and academia?

The foundation is code. At heart, the novelty in the work of data scientists is not novelty at all - it is just the discovery that being able to use code makes a huge difference to the analyses that you can do, and the conclusions that you can draw. Analysis based on code is:

Versatile. Code allows us to analyze big, messy, mixed, and complex data - put more simply, it allows us to analyze real data.
Realistic. The real work of analysis is working with real data. Much of this work is cleaning, reorganizing, collating and exploring. It is both complex and repetitive; it can only be done effectively with code. If you can code, then you can do this work, describe it, and teach it.
Collaborative. Programmers have been working for many years on methods for effective collaboration. Data scientists learn these methods, and apply them, so they become more transparent, more efficient, and better at working together in teams.
Reproducible. Much analysis that does not use code, uses graphical interfaces. These are famous for making it difficult to record your analysis, or describe it to someone else. Analyses based on code are naturally reproducible, because you can run the code again, to produce the same result, and you can give someone your code, so they can do the same thing.
Easier to understand. Coding allows us to shift our emphasis from the relatively difficult mathematics behind the old-school statistical tests, such as t-test and ANOVA, to more direct methods such as resampling, including permutation tests and bootstrap estimation. Code gives us a language to describe these methods in a simple and direct way, and this makes the ideas clearer and easier to generalize.

Why now?¶

Donoho and others have pointed out that much of what is distinctive in the new culture of data analysis was already part of the analysis described by John W. Tukey, in the mid-20th century. Why have these methods only now started to take hold? We believe it is largely to do with the advances in the tools of scientific computing. Over the last 20 years we have seen the growth of a new generation of programming languages with clear, powerful syntax, such as Python. Another language widely used in data science is R, the statistical programming language. The originators of R were thinking specifically of data analysis when designing the language. Python and R are free, and open source, so they have benefited from the explosion of productivity in open-source software. As these languages have developed, they have attracted more scientist-programmers, who build new libraries for data analysis and statistics. Meanwhile programmers in industry have learned the hard way, how difficult it can be to write clear and effective code, and there is a substantial body of thought and practice on process for writing and sharing code. This process spread into the world of open-source programming, leading to a great increase in the quality and efficiency of academic code. The combination of richer, clearer languages, better libraries, and better working process have made it possible to do much more with code than was previously possible. As a result, many more scientists can write good code, and therefore, many more scientists use code for data analysis. Eventually, these scientists appeared in companies that had difficult problems in data analysis, and showed how effective these skills could be. These were the data scientists described by Patil and Hammerbacher.

Elements of data science¶

If we are right, then data science is defined by what data scientists do. What do do they do?

They collate data from many sources;
They clean up messy data, and use it to try to recover meaningful information. But they are sceptical, and know when the data cannot be trusted for strong conclusions.
They explore data, to find errors, and unexpected patterns. They will likely use different ways to make graphs and graphics from the data, to diagnose and summarize.
They work hard to understand the process that generated the data, to make their conclusions meaningful.
They understand variation, and the problems and limitations for drawing conclusions from noisy and incomplete data.
They often try and predict what new data will look like, from old data. They will use old and new methods to do this, including methods from machine learning and statistics.

Conclusion¶

Data science is new, and we are still working out what it is. At the moment, it is best defined by what a data scientist does. A data scientist uses programming as the basis for a deeper, more flexible approach to data analysis.

References¶

Bre01: Leo Breiman. Statistical modeling: the two cultures (with comments and a rejoinder by the author). Statistical science, 16(3):199–231, 2001. URL: https://projecteuclid.org/euclid.ss/1009213726.
DP12: Thomas H Davenport and DJ Patil. Data scientist: the sexiest job of the 21st century. Harvard business review, 90(10):70–76, 2012. URL: https://hbr.org/2012/10/data-scientist-the-sexiest-job-of-the-21st-century.
Don15: David Donoho. 50 years of data science. In Princeton NJ, Tukey Centennial Workshop. 2015. URL: http://courses.csail.mit.edu/18.337/2015/docs/50YearsDataScience.pdf.
NAoSM18: National Academies of Sciences, Engineering and Medicine. Data Science for Undergraduates: Opportunities and Options. The National Academies Press, Washington, DC, 2018. ISBN 978-0-309-47559-4. URL: https://www.nap.edu/catalog/25104/data-science-for-undergraduates-opportunities-and-options, doi:10.17226/25104.
Pre13: Gil Press. A very short history of data science. https://www.forbes.com/sites/gilpress/2013/05/28/a-very-short-history-of-data-science, 2013. URL: https://www.forbes.com/sites/gilpress/2013/05/28/a-very-short-history-of-data-science.
Tuk62(1,2): John W Tukey. The future of data analysis. The annals of mathematical statistics, 33(1):1–67, 1962. URL: http://projecteuclid.org/euclid.aoms/1177704711.

Coding for Data - 2020 edition