How Data Science Teams Can Win By Behaving Like Lean Startups

Ben Dias Royal Mail’s Head of Data Science reveals a practical guide to acting like a startup

23Mar

There are many ways in which data science teams are like startups. By adopting a lean startup methodology, corporate data teams can navigate internal politics and deliver bigger and better results.

Speaking at Predictive Analytics Innovation Summit, Ben Dias shared for the first time in public how Eric Ries' seminal book on startups has helped him to achieve successes in building high-performing data science teams at Royal Mail and Tesco.

In his first year at Royal Mail, Dias and his 20-strong data science team have landed 5 new projects worth £50 million. They have also launched a number of transformational prototypes and killed off 15 legacy projects.

What’s more, the board are so pleased with the results so far that he has been asked to double the size of the team. These successes come in no small part from how Dias has adapted the lean startup philosophy to data science.

"Data science should be an entrepreneurial venture, something that a company hasn’t done before. We try to solve problems where a solution is not obvious and success is not guaranteed, which makes data science very similar to startups."

Dias further explained his approach:

Innovation accounting framework

"The biggest mistake data scientists make is jumping straight into a solution and wasting months doing the wrong thing. This is not lean."

Spend time with stakeholders to define the definition of "done", "best" or, "better"?

Tesco asked for an algorithm for "best range" but couldn’t answer exactly what "best" meant.

If we take one aisle of a store and optimize for sales value, the aisle becomes full of beer which doesn’t present a good customer choice.

Stakeholders can be better at telling you what isn’t right than what is!'

Then, before you jump in, figure out the Innovation metric that helps you to track your progress and communicate to stakeholders. Success should be judged by one metric (or at most, no more than three) must be easily understood.

"Metric is a mix between science and art'” says Dias, advising that it is helpful to 'Think of metrics as people. If you can track progress by the number of people we affect e.g. 50% of customers did this, you won’t go far wrong."

"Then, brainstorm for a Minimal Viable Product (MVP).

What is the simplest way of tackling this problem? Start with a logistics regression first before AI and see if it's good enough. If it is already 80% accurate, then that could do, so move on to the next project.

'The key to failing fast is tracking your innovation metric. If going for 2-3 cycles and metric isn’t moving, stop and think about it.

Pivot or persevere. It could mean killing a project, bringing in different data sets or completely changing the project.

The key to not spending six months on a project which isn't working is tracking its innovation metric and killing it fast.

Be agile about agile

Dias uses both scrub and Kansan project management methodologies.

"Do whatever is right for the thing you’re trying to do. Scrum is predefined work for a set time, it assumes you know what to do. While Kanban is better for dealing with uncertainty, such as dealing with bugs once a product has been deployed."

The team at Royal Mail use Jira with four boards and use either scrum or Kanban

  • MVP (scum board)
  • Development (Kanban board)
  • Deployment (Scrum board)
  • Support (Kanban board)

For more on the perfect data science sprint by Kat James at Royal Mail

Hypothesis-driven approach

Break the problem down into key components and risk assess first, prioritize and test.

'If getting a particular data source is the biggest risk, chase the data first. If you can’t get data in a couple of weeks, kill the project. Don’t waste any time until the riskiest thing is sorted.'

Iteratively experiment with everything

'You can even apply it to the recruitment process. "Looking for 20 people this year."

Retrospectives are key

'Pull the whole team together to talk about what went well and what can improve. Come up with only 3 actions based on feedback. As a team, we get feedback and learn.'

A flat structure where everyone is empowered

While it’s impossible to line manage everyone, Dias makes clear that management is for pastoral care. When it comes to a project, everyone has an equal seat at table compared to seniors. 'I expect data scientists to challenge everyone including me. If I can’t explain it well enough then don’t do it.'

In addition, both the data science and data engineering teams now report into Dias, a change which has made it easier to get results.

How to kill projects quickly

One of the greatest time savings that this approach offers is making it easier to kill projects.

'Legacy projects are much harder to kill. Almost took a year to kill some of the projects I inherited.

But new ones have total control. The key is to engage stakeholders with a lean startup approach

Make sure you understand Fail Fast and that when we take on any project, we not committing to finishing the whole project, we are committing to 2-4 weeks before deciding if we’re going to kill the project and then 6 weeks deliver MVP.

Involve them in the pivot meeting to see what we do next. Take them on the journey with you.'

The next data analytics event is the Big Data Innovation Summit in San Francisco.

Dataviz

Read next:

Why We Need Data Visualization To Understand Unstructured Data

i