Finally Some Sense on Analytics & Data Science Job Ads

After yesterday's post on the state of the debate on building data science teams (individual vs team approach), it's so refreshing to stumble onto this careers page of Civis Analytics. Great example of Analytics & data science job ads done right. This page alone makes me want to apply to work there!

They actually divide their jobs into: data scientist, engagement analyst, project manager, software engineer!

How sensible. I like it! 

Nothing like the typical data science job posts, asking for "everything and the kitchen sink".

Building Data Science Teams: Individuals vs Team - State of the Debate So Far

Since my last article on "Hiring 1 Data Science unicorn is hard enough, a team is impossible. To scale means to specialise", similar ideas have been expressed by InformationWeek, McKinsey/HBR, and KDnuggets (here, here, here and here).

There has a been a ton of great discussion. I attempt to summarise the viewpoints so far: 

• Data Scientists are supposed to have some pretty deep expertise in some pretty hard areas (see diagram). 
• Is it possible to close this talent gap when we seem to be chasing after superheroes or unicorns? (there are some, but very few)
• Some (44%) think there should be data science sub-specialisations (which all exist today), and have them work together in a team.
• Others (44% too) prefer the superhero approach - individuals who have it all

Opinions so far on the approach of team vs individuals to build out a data science team are as follows:

For Team / against individuals	For Individuals / against team
for bigger companies	for smaller companies (can't afford)
Easier to find all necessary skill-sets	Easier to get things done (no coordination friction)
Don't fall apart if an individual leaves
Jack-of-all trades, master of none; Deep expertise more possible in team	Automation tools will take over data engineering & cleaning from DS jobs, so can concentrate on modelling
Business domain expertise & soft skills are hard to find in math/quant majors	Higher-ed will turn out DS superstars soon, who will have the combined maths/computing skills
A good team has both Specialists and Generalists
DS is a field that's evolving fast, and so will these opinions
You want an all-round DS guy/gal to get you started, or 2-3 of them who round each other off. As your team grows with demand, it will become increasingly difficult to find those all-encompassing individuals, so your team will naturally be people with 1-2 of the DS skills.

If you are still keen to know more about what data scientists do, and who they are, listen to these DS guys talk:

• Amazon's principal engineer: John Rauser, "What is a career in big data?" - 17 minutes of a very good stepped-back view of data science.
• Cloudera's director of data science: Josh Wills, "Life as a data scientist" - some good nuggets in there at minute 10, 16, 25, 52:
• "I'm a competent statistician... I'm a competent programmer... I would not say I am good... I am capable of having a conversation at each of those fields with them..."
• "Scientists get linear regression...but they don't get the difference between linear regression and logistic regression...or the assumptions that underlie the regression models", like normal distribution of the variables for linear regression; it's more of a "mechanical" exercise to turn the crank on the data without understanding the assumptions that support the model
• Kaggle has "done most of the hard work [for the competitors]". In my opinion, the guys who are competing are good at using the ML tools on a clean'ish data set; but it doesn't exactly test their ability to go from a business problem to a "mental model of the data required" to the type of problem to solve (segmentation, regression, etc...)
• what stats to learn for someone from the computer engineering side of data science: "learn linear regression, t-tests, confidence intervals, binomial random variables, exponentially distributed random variables, ... the core stuff, really, really well"

P.S. After writing this all out, it sounds so obvious. But believe me, there has been so much debate around this topic, and I wanted some... sense. Go read those articles linked at the top if you want to know.

Hiring 1 Data Science unicorn is hard enough, a team is impossible. To scale means to specialise.

The Data Scientists need a large set of skills, including business know-how, modelling and mathematics, plus programming. They are as hard to find as unicorns, or superheroes. I know this talent shortage first hand. Is the solution to create more unicorns, or can we devise better solutions?

In my last role as a managing consultant in the Operations Research and Analytics team of a large global consultancy, I also ran recruitment. Having spoken to or met 150-200 of such candidates personally, and my recruitment team saw multiples of this number, I can tell you not many of those candidates made the cut. That's because they didn't have all of the skills we were looking for. And we were only looking for the first 2.5 of the 5 core skill-sets of a data scientist below. "Good luck" is what people offer to this talent-search problem, but I think we can get around the unicorns.

The 5 core skills of a Data Scientist 

Expanding on the data science venn diagram, I think the following 5 skills deserve closer attention, separately*.

• Business consulting (from problem definition to stakeholder and team management) :: what problem to solve
• Analysis and modelling (maths, stats, physics, OR, engineering, etc. / note this includes coding) :: how to solve it
• Communication and visualisation (artistic and functional, learn the visualisation tools) :: how to tell the story
• Data engineering (take data in, store it, push data out: computer science) :: how to get the data for the solution
• Programming (for enterprise use at production level, software engineering, integrating into BI systems, automated decision making embedded in operational systems) :: how to make the solution useful to a wider audience

Furthermore, each of the above have subfields and specialties, because they are complicated in their own right. It is not possible to be very good at so many things, not at scale anyway or to be above mediocracy at best. How many sportsman/woman excel at more than one sport, for example?

It's a lot to ask for one person. So, why ask just one person?

The thing is, these people all exist, have existed, and will exist. They are just separate individuals. They have labels like business analytics consultants, statisticians and modellers (operations researchers included), data visualisation experts, DBAs and software engineers. Yes, they are also talents in need, but they are not unicorns. If we need data scientists in troves, we need a team, not just a few geniuses.

The future I see is like the age old relationship advice: 

Don't try to change them. Instead, let's change how we work with them.

People should diversify a bit, for instance a modeller should be able to code, but ultimately they need to specialise in something they are good at. A modeller must be able to prototype on his/her own, which requires coding skills, but s/he shouldn't be expected to produce production-ready code for large scale applications. Similarly, asking a good modeller to do database administration and ETL tasks is a waste of talent, Hadoop or not.

Specialisation is the reason for humanity's proliferation. Therefore, I'd say it's not the people we need to change, but the system that we need to setup to allow such specialised workforce to team up together. It's lazy for the analytics field to put up its feet and just summon one person to provide it all.

As a starter-for-ten, I think the future of our field could be modelled after the traditional IT project group make-up:

• The technical "purists": analytics modeller, data engineer, visualiser, programmer
• The "bridge": more like a traditional business analyst
• The "glue": project manager with business consulting skills

There will be complications to address. To name a few...topics for another post:

• Who should start up your data science team?
• What's the load balance? (how much of each skill to have)
• How to coordinate the division of labour?
• Where should they sit in the organisation?
• How to prioritise the problems to set them working on?
• How to manage this team?
• What's the career path?

Where do you think the analytics field is heading to?


Disclaimer:
My views are definitely biased by my background: I am a manager in business analytics consulting, trained in Operations Research and Computer Science.

* My expansion on the data science venn diagram's 3 skills are based on various articles, such as O'Reilly guide, Intro to DS skills, and job requirements on numerous current data science job posts.

From Operations Research to Data Science

In the last post, I wrote about how good it is to see OR linked as a skillset to data science. However, do note that OR is only one part of the DS skillsets. OR ≠ Data Science. How does an Operations Researcher transition to a Data Scientist?

There are a few things the O'Reilly book I talked about in the last post briefly mentions as suggestions for an OR person to learn more about: some of the new Bayesian / Monte Carlo Statistics methods, broad programming skills, data warehouse architecture for big data technology, and business kills "to be able to intelligently collaborate with (or lead) others on a data science team".


For those looking to upgrade, here are my quick thoughts on where to start.

Bayesian Data Analysis: Andrew Gelman from Columbia is running a course on Bayesian Data Analysis *right now*, with Google+ Hangout sessions. Looks very interesting.

Programming skills: see my previous post on learning R and Python - the languages of data science.

Big data architecture: in my experience, first understand the layers of a normal data warehouse architecture, then broaden to the enterprise BI architecture stack, then learn about the new bits for addressing the "big" aspect. I was fortunate to have led a fairly big project in this area, and had the opportunity to work with some great data warehouse architects and enterprise BI architects to learn a ton from them. I'm not sure what the best self-learning material is other than the typical read-a-lot. Wikipedia doesn't seem to cut it, and the best material that helped me aren't publicly available. Hmm...I will have to think about this - topic for another post perhaps. In the meanwhile, Pivotal seems to do a fairly good job in their blog to dumb down the explanation of the bits for "big" data technology in some practical terms.

Business skills: I think this only applies to academics (sorry for the generalisation). For the practitioners, i.e. OR people working in and with businesses, that's a fundamental part of our jobs.

Operations Research is a skillset of a Data Scientist

... according to O'Reilly, yes, it is. 

This is perhaps the clearest I've seen anyone link OR to Data Science. Or perhaps, depending on how you read it...it shows that some Data Scientists are OR people. OR is a subset of Data Science skills.

Data Scientist (DS) - a very popular label that seems to be associated with people kind of like us-OR-people these days (just like "analytics" has been for the last few years), but no one is completely sure exactly what it is. As a result, many of us are reluctant to call ourselves a data scientist, or don't know how to make the transition to be called one (see my next post on where to start). There is the Venn diagram, and examples from famous DS people like Nate Silver and Hilary Mason (who are identified more as statisticians than anything else), but confusions are still bountiful.

OR has always had a bit of an identity crisis - how many jobs have you seen with the words "operations research" in the title or description? Is "Data Science" here to help?

O'Reilly published a book, titled "Analyzing the Analyzers", which discusses the results and implications for people in these related fields, based on a survey they ran in mid-2012, with whom they consider as data scientists, and "how they viewed their skills, careers, and experiences with prospective employers". Their goal, best summarised in their own words, are, "in the broad Analytics / Data Science / Big Data / Applied Stats / Machine Learning space, ...to define these new fields better, and we hope the results will help people such as yourself talk about how your skills and your work fit in with everyone else's."

The main result was summarised into a 5X4 matrix (credit: O'Reilly), showing where the survey respondents are in terms of skills / expertise and the label they associate themselves with. 

The list of skills they grouped under "Math / OR" are: Optimisation, Math, Graphical Models, Bayesian / Monte Carlo Statistics, Algorithms, Simulation. Sounds familiar indeed.

Hooray for the mention of OR as a Data Science skillset!


I recommend reading the full report for more details. Here is a summary to give you a taste:

• Four data scientist clusters
• Cases in miscommunication between data scientists and organisations looking to hire
• Why "T-shaped" data scientists have an advantage in breadth and depth of skills
• How organisations can apply the survey results to identify, train, integrate, team up, and promote data scientists

(The last point above: it wasn't too comprehensive, so don't expect too much. More of a taster.)

Have you got what it takes to call yourself a Data Scientist? OR folks, see my next post on how to upgrade yourself (umm, didn't mean to make you sound like machines).

How to Learn Python and R, the Data Science Programming Languages, from Beginner to Intermediate and Advanced

The Data Science programming / analytics languages to know are, R and Python. If you're in Operations Research or another analytics field that somewhat fits under the "Data Science" hat, you: a) already know them really well, b) want to brush up on them, or c) you probably should learn them now. Here I compile my thinking on how to learn R and Python from Beginner to the Intermediate and Advanced levels, based on having tried some of these course materials.

Beginner (doing basic analysis)

R:

Computing for Data Analysis on Coursera and Youtube (weeks 1, 2, 3, 4), by Roger Peng from Johns Hopkins University

• Summary: It covers the basics of conditioning and loop structures, R's syntax, debugging, Object Oriented Programming, performing basic tasks with R, such as importing data, basic statistical analysis, plotting and regular expressions. See syllabus for more.
• Time commitment: 11~36 hours total, including: 
• non-programmers: 4 weeks X [3 hours/week on video + 2~6 hours/week on exercises]
• programmers: [3 hours of notes reading + 8~16 hours] on exercises
• Advice for: 
• non-programmers: Listen to all lectures (videos), make sure you understand all details, and do all the exercises to hone your skills. Programming is all about practicing. Doing the exercises are important. See below for "Advanced".
• programmers: Don't bother with the videos, go straight to the lecture notes (link). Read the notes - much faster than the videos. if you don't understand anything, look up the video and watch, or google the topic. Then do all the exercises. You don't need me to tell you that practice is king (um, and cash too).

The swirl package within R, by the Biostatistics team at Johns Hopkins University

• Summary: It aims to teach R and Statistics within the R environment itself, through a package called swirl. See the announcement here for more detailed info.
• I haven't tried this, so I'm not sure how much time it takes or how good it is. However, I think it sounds pretty good, and deserves a mention. I was never a fan of reading books to learn a programming language. Show me the code, or in this case, let me write the code, and get involved, is much more, well, involving.

Python:

Google's Python course (link)

• Summary: It's straight to the meat, no non-sense stuff, and covers all the important things. Suits my style. Enough said, so see the course page on the syllabus. 
• Time commitment: 8-10 hours
• including reading notes and doing exercises
• Note, this is for experienced programmers. There are videos too, but don't bother. The notes on the course page are the same, and it always takes less time to read than watch.

Intermediate (building analytical models)

R:

Data Analysis with R on Coursera and Youtube (plus class notes), by Jeff Leek from Johns Hopkins University

• Summary: It covers the full modelling cycle, from getting data, to structuring the analysis pipeline, exploring with graphs and statistical analysis, modelling (clustering, regression and trees), and model checking with simulation. It also talks about important statistical watch-outs like p-values, confidence intervals, multiple testing and bootstrapping. More syllabus here.
• Time commitment: 32~56 hours
• including 8 weeks X [2~3 hours/week videos + 2~4 hours/week exercises]

Forecasting using R (link), by Rob Hyndman from Monash University in Australia and Revolution Analytics (the enterprise R solution)

• Summary: topics include "seasonality and trends, exponential smoothing, ARIMA modelling, dynamic regression and state space models, as well as forecast accuracy methods and forecast evaluation techniques such as cross-validation. Some recent developments in each of these areas will be explored" (quoted from course site). Read more there.
• Note: I haven't done this (just started), so I'm not sure about its time requirement or quality. I'm also not sure if they are planning to make available the lectures. Time will tell on these questions.

Python / Octave:

Machine Learning on Coursera, by Andrew Ng from Stanford University --> My Favourite!

• Summary: The course actually teaches in the Octave language, but it all can be done in Python. I suppose you can do it twice, first in Octave, and then in Python, if you've got the time. It certainly would solidify your understanding of the material, and Andrew Ng is sure that Octave is rather important in Machine Learning. It assumes some prior knowledge of linear algebra and probability, and refreshes you on some basics. "Topics include: (i) Supervised learning (parametric/non-parametric algorithms, support vector machines, kernels, neural networks). (ii) Unsupervised learning (clustering, dimensionality reduction, recommender systems, deep learning). (iii) Best practices in machine learning (bias/variance theory; innovation process in machine learning and AI)." (quoted from the course website)
• Time commitment: 50~90 hours
• including 10 weeks X [2~3 hours/week videos + 3~6 hours/week exercises]
• Note: this course covers a subset of the statistical and modelling principles from the Data Analysis with R course above, but the overall level is more advanced. I enjoyed this course the most.

Advanced (you follow the drift from above)

Advanced = Experienced.
This is true for programming, analytics, and learning any foreign languages.

"Just do it", is how you get experienced.

There is no course on this stuff (i.e. being advanced), not without a PhD _plus_ years of field work.

My best suggestion is use your curiosity. Find a problem. Dig into it.

Plus, work with other people that are really good.

Happy learning!

More MOOC on Analytics - Coursera

A hoard of analytics related Massive Open Online Courses (MOOCs) are about to start in September. Have your pick on what to learn. Having taken a few Coursera courses now, I would recommend 1) not taking too many courses at once, however tempting it is to sign up to all of them, unless you have no other work or projects on the go. This is just to make sure you have a reasonable load and able to devote enough of your attention to learning the material properly. 2) Make good use of the discussion forums, as they are both a good source of clarifications and a window into other people's perspectives on the material. 3) Do the exercises, programming assignments and quizzes to ensure your understanding of the material.

Linear and Integer Programming
Starts 2 Sept 2013, 9 weeks, 5-7 hours/week
(the basics of mathematical optimisation, a core toolkit in the field of Operations Research)

Statistics One
Starts 22 Sept 2013, 12 weeks, 5-8 hours/week

Introduction to Recommender Systems
Starts 3 Sept 2013, 14 weeks, 4-10 hours/week

Computing for Data Analysis
Starts 23 Sept 2013, 4 weeks, 3-5 hours/week
As I've written before here.

Web Intelligence and Big Data
Starts 26 Aug 2013, 12 weeks, 3-4 hours/week

Thinking Again: How to Reason and Argue
Starts 26 Aug 2013, 12 weeks, 5-6 hours/week
Perhaps a bit off topic, but perhaps not, since all analytics are more or less rooted in proving or disproving arguments, so we better learn how to do it well.


Related article:
Coursera and the Analytics Talent Gap
Starting up in Operational Research: What Programming Languages Should I Learn?

Everybody likes to predict, but nobody likes being predictable, nor told what to do

The Netflix algorithm is in the news again.
The Science Behind the Netflix Algorithms That Decide What You’ll Watch Next

Netflix finds rating predictions are no longer as important, trumped by current viewing behaviour, i.e. what you are watching now. However, browsing through the comments, and again, you will see a generally negative reaction. Some people really hate being told what to watch, even if it's just a recommendation. Others say Netflix sucks, because it recommends things they've watched elsewhere. That sounds like a lack of understanding: if you don't tell Netflix you've watched something already, then how could it know?

As "big data" gets more media attention, it is reaching a wider audience who don't yet understand how algorithms work, but only know there are algorithms everywhere in their life, and it's scary to them. The lack of understanding seems to create fear and resentment.

LinkedIn and Facebook's recommendation systems for helping people find colleagues or friends they may know are generally well received, yet these film recommendation systems aren't. The difference between them might underline the success criteria of rolling out such recommendation systems.

Machine Learning in Movie Script Analysis Rouses Angry Reactions

An application of Machine Learning is covered in the news lately: movie script analysis.
Solving Equation of a Hit Film Script, With Data

They "compare the story structure and genre of a draft script with those of released movies, looking for clues to box-office success". However, the comments reveal that the general population (at least of the commenters) dislikes the concept for fear of anti-creativity.

Comments like these sum up the overall sentiment:

"Using old data to presage a current idea is both terrible and foolish. It is to writing what Denny's is to fine dining - mediocrity run wild."   

"Data crunchers will take the art out of everything. Paint-by-numbers."  

Ouch.
You be the judge whether this is a good application or not.

I tend to bias towards answers like this from the comments (sadly this was only 1 of 2 positive comments at the time of my reading; the other one was from the CEO of the script analysis business):

"I'm sure people have all sots of assumptions about what audiences like already. This data could be a tool to look deeper into these assumptions. Film makers have always wondered about consumer taste. It is a business. When commerce and art mix, there are inevitable compromises. This tool helps people see possible preferences based on past behavior. Information should never frighten us. It is how this information is applied that most deserves our attention." 

I think it also never helps the image of such machine learning practitioners when the journalist tries to paint him with an antagonist brush, such as "chain-smoking" and "taking a chug of Diet Dr Pepper followed by a gulp of Diet Coke and a drag on a Camel". Reminded me somewhat of another writer's writing style when covering analytics.

Our labels: data scientist vs statisticians (or OR)

A perennial discussion of identities in the world of analytics is making the rounds on the blogs of statisticians. Or wait a second, what should we call them?
Data scientist is just a sexed up word for statistician

Data Scientists, Statisticians, Applied Mathematicians, Operational Researchers...jus to name a few, are the labels one might apply to themselves in the field of analytics. How shall we label ourselves? I can't agree more with Nate Silver,

"Just do good work and call yourself whatever you want."

Learn R with Coursera for Data Analysis

Heads up: the Computing for Data Analysis course is running in September 2013.

It will teach you the R language for data analysis. The course is described as:

This course is about learning the fundamental computing skills necessary for effective data analysis. You will learn to program in R and to use R for reading data, writing functions, making informative graphs, and applying modern statistical methods. 

In this course you will learn how to program in R and how to use R for effective data analysis. You will learn how to install and configure software necessary for a statistical programming environment, discuss generic programming language concepts as they are implemented in a high-level statistical language. The course covers practical issues in statistical computing which includes programming in R, reading data into R, creating informative data graphics, accessing R packages, creating R packages with documentation, writing R functions, debugging, and organizing and commenting R code. Topics in statistical data analysis and optimization will provide working examples.

Related article:
Coursera and the Analytics Talent Gap
Starting up in Operational Research: What Programming Languages Should I Learn?

Coursera and the analytics talent gap

It's been a while, and ThinkOR is back to blogging about Operational Research and its related themes.

ThinkOR authors are about to start on 3 Coursera courses over the next couple months:

• Natural Language Processing
• Machine Learning
• Data Analysis (with R) (to refresh)

I am not only learning about some new topics for my own benefit, but also interested in assessing how such easily accessible courses could help the so-called 'big data and analytics talent gap' in businesses. As a Business Analytics consultant, this is one of the biggest issues I see my clients facing in today's business world - one wouldn't think about it, if they don't know about it, and once they know about it, they don't know how to get more of it. Obviously, there would need to be some sort of a step progression, such as (just an example without much research at this point):

1. Statistics One
2. Data Analysis (with R) and/or Computing for Data Analysis
3. some sort of programming course, check the computing course catalogue
4. Focus on one or several of the main OR techniques and their associated tools, such as Discrete Event Simulation, Monte Carlo Simulation, Optimisation, Forecasting, Machine Learning, and the good old Volumetric Modelling, as some examples
5. and if you are going to work with humongous data sets, Intro to Data Science sounds reasonable to become familiar with the various big data technology to apply data science (I suspect this often eludes traditional OR practitioners)

As ThinkOR goes along, we will be blogging about these courses and our learning experience. So far, there has only been very positive feedback. Let's get going!

Merry Christmas and Happy New Year!