spencer nelson

Data science is about company culture

You can find gobs of chatter out there about what makes a good data scientist. Most focus on technical aspects: know statistics, write scripts, build machine-learning tools. Sure, these are all important, but the list is incomplete. The most valuable skill for a data scientist is the ability to shape the culture around them, to convince others to use data to inform their decisions.

The best data scientists are the ones who transform their company. They motivate an entire organization to make decisions by measurement and numbers. They do not do all analysis themselves - they turn those around them into analysts.

People who have machine learning skills, know statistics, and can write code are terrific, yes. But those alone are not transformative.

why do companies want data scientists?

When companies hire data scientists, they rarely want a wizard who can dazzle with highly technical presentations but who works alone. Impressive algorithms and statistical depth are not the mission. The true goal is just to make better-informed decisions. The best way to do that is to make data-informed decisions routine to the organization.

how does this happen?

There are two components in transforming a culture. You have to convince people that they should want to work differently, and then you have to provide the tools to make that possible.

These aren't really sequential steps - they're tied tightly together and happen at the same time.

convincing people

How do you convince people to work differently? To some extent, this can be done with dazzle: impressively prescient predictions can convince most people. But the risk is that people think data science is a technical task which they can't accomplish, which is really the worst possible outcome. I think that the right answer is to always jump into decision-making discussions and constantly push for metrics to back up claims. Good data science requires constant engagement with people making decisions, making a quantitative perspective a natural part of the whole process.

giving them tools

How do you provide tools to get people to use data? The buzzword here is 'data democratization'. Dashboards and non-technical displays are really core to a data scientist's skillset. I don't mean infographics - I actually mean the opposite. No-nonsense, straightforward tables that provide access to clean data for anyone in the company can be the single most valuable internal tool you will ever have.

I am not exaggerating. This is deceptively hard - clean data can be hard to come by, and presenting data in a straightforward way - while capturing the nuances and quirks which are inevitable in the real world - is very difficult. But the reward can be a culture in which 'checking the numbers' is a routine part of any workflow, from customer service to devops to finance.

how to do this

If you're a data scientist, here's my advice:

  1. Make yourself incredibly accessible. Offer to help people get data whenever they need it, and make it a priority to get it to them fast. The speed is key - it will help make the data process seamless, which will make it easier to convert it into a habit.

  2. Make dashboards and web interfaces to your scripts like crazy. I use Flask for this purpose, wrapping tasks up into small applications. Build them aggressively - any time you're running a script twice in a row is a sign that this should have its own web-based control system to open it up.

  3. Try to be a part of all planning meetings you can manage. This will be a little painful, but it's important: you want to be there when people make decisions to immediately prod them on whether they have thought about data. You don't need to talk much in these meetings - you might offer to just take notes. As a bonus, you'll gain a lot of the contextual knowledge which is so crucial for modeling.

If you're an organization with a data scientist, encourage them to do this stuff. The important thing is to make sure that data becomes a culture, and that data science is not a silo within the company doing abstruse analysis, far from other people. You want data to be integrated at every level.

It's not as sexy as visualization, and it's not as intellectually intimidating as machine learning, but it ends up being worth a lot more.

Discussion at Hacker News: link