So, we have a serious problem in Data Science at the moment. We have been doing Data Science for a few years now, anecdotally we know many AI projects fail, companies are struggling to retain data science managers and data scientists and the project sponsors of Data Science initiatives are starting to ask questions.
In our excitement about Data Science and AI it seems we put together teams of Data Scientists, flicked them some data and said “Go for it guys!” without the right structures, framework and strategy in place for them to succeed. What we need is a bit of a reset and a process for executing data science projects with the right people in place to
execute these projects.
Fluffy Corporate Strategy
It all starts with a fluffy corporate strategy that is going to be by definition too high level for the Data Science team. Most will be along the lines of making something better.
This kind of corporate vision just doesn’t map well to Data Science Teams. Data Science Teams need numbers and metrics, we need to tease out a KPI or a hero metric for the Data Science Team and timelines to execute. We need to give them something to pitch for, something to optimize. Otherwise the Data Science team will not be aligned strategically to the corporate vision of the company. This is where the Data Science Managers manage up to the execs and managing down to the members of the Data Science team.
Mapping the Fluffy Corporate Strategy to the Data Science Team
Let’s look at a concrete example below.
So we have our high level strategy of lets “build long term customer relationships”. As we know this just won’t work as an objective for our Data Science Team. So we need to map this corporate strategy to something the Data Science Team can aim for. Let’s say something more meaningful would be:
“Let’s reduce customer churn from 10% to 5% in the next year”
Now that’s something we can work with!
We may do some more analysis and discover that a great proxy for churn is if someone has no interaction with website in 90 days the customer may then have a 95% chance of never coming back to us! Now we have this proxy we can use it in our data analysis, our reporting dashboards, our KPIs, we can then set up A/B tests.
Heck we can even grab past data, attach our outcome to that data and build a churn model to run monthly and then experiment with interventions such as sales calls, discounts etc.
Technical Leaders Required for Strategy Execution
Steve Jobs tried to hire professional managers in the early days of Apple, but he found there was a problem with doing this, they didn’t know how to do anything:
“We went out and hired a bunch of professional management. It didn’t work at all. Most of them were bozos.”
The best managers were “the great individual contributors” who didn’t necessarily want to manage, but fill that role simply because they are the best person. In other words he created technical leaders, and so should you!
Great people in Data Science aren’t going to want to be managed by someone who isn’t technical, who isn’t the sort of person they can learn from. Even if they are paid well, eventually they will find another job.
So, a misaligned strategy and a lack of technical leaders is why Data Science projects fail before they even get off the ground
A Process for Reliably and Consistently Executing Data Science Projects
We have seen this problem before, but it was in software development in the 1990’s, this actually led to the Agile Manifesto. Agile software development was a paradigm shift compared to the way software products were built and delivered. Simply put you’d plan maybe a year or even longer in advance at the start, and then work for a year and deliver something. If rework was required at any stage you’d have to go back to the drawing board and start again. So, the problem was most software development project were running over time and over budget.
The Agile movement emphasised MVPs, short development cycles of 2-3 weeks, just as much documentation as was needed and regular feedback from the customer.
We have an equivalent now for Data Science, the DataOps Manifesto:
The Microsoft Team Data Science Process sets out:
A framework for executing Data Science projects
The roles and responsibilities for those involved
Template artefacts, code and folder structures
Implementation guidelines and how-tos
It is absolutely comprehensive, and in my mind is the way to manage Data Science Teams and to go from idea, to data acquisition, to modelling, to ROI enhancing data product in a production system. It is simply brilliant. One caveat is that it requires technical leaders to execute.
Hire the Right People at the Right Time
What I have seen in my time working with the big end of town all the way to little startups with 4 people is that there is a spectrum of analytics sophistication. Just like aspiring data scientists organisations are also on their own Data Science journey.
So don’t make the mistake of hiring 20 PhDs to automate a spreadsheet. You might need a technical BA and a BI developer initially. Then go ahead and hire a DBA to set up databases for the data scientists to work with. Then you might go about hiring data scientists when the time is right on an “as needs” basis.
Beware of the “Icing on the Cake” Problem
MOOCs and universities sadly don’t tend to focus their education of Data Scientists on the day to day, the housekeeping of Data Science. That is the data munging, preparing datasets, exploratory data analysis, the hundreds of little decisions that go into creating a dataset suitable for model development.
Instead they tend to focus on what I call the “icing on the cake” which is just the statistical modelling part.
Real data is messy, real Data Science problems are opaque. You need to hire for people who have real world experience and are comfortable dealing with problems that are not clearly defined, or if you are going to hire recent graduates (and I think you should!) look to those who have competed end to end Data Science projects like the kind you would see in your organisation. Not just the same old tutorials everyone else is doing. If you are a student or recent graduate focus on your own end to end Data Science projects.
Business Acumen and Communication Skills (the Secret Sauce for Data Science Success)
There are a couple of “soft skills” you’d really want to tease out in candidates if you can.
If you have a commercially focussed Data Scientist who can communicate both verbally and in writing then you have picked a winner.
In an environment when just about anyone can claim to be a Data Scientist, and in many respects anyone can over time develop the technical skills required to do the job, these two skills stand out as being huge ways of differentiating the top talent.
Democratize Data Across the Organisation
Once your Data Science team has had some wins, it’s time to make Data Science more than just about the Data Science Team. It’s time to embed members of the Data Science Team within the different business units. Have them attend the meetings of the team, discuss the way data can be used for decisions and then feedback revenue making opportunities to Data Science Managers to start the cycle of making ROI again.
Rinse and repeat to continue to add value!
Create a Data Driven Culture from the Top Down
If you are able to execute this strategy, you can significantly reduce the chance that your Data Science projects will fail. You will better be able to retain your Data Scientists and Data Science Managers and you will have success where your competitors won’t. What’s more I think you will create a collegiate environment across your business and just have a heap more fun.
Isn’t that what this is really all about?
All the best, and if you have questions feel free to reach out.