Failing Your Way to Success: Embracing Experimentation As a Data Science Leader
If there's one thing that sets data science projects apart in the broader IT or "technical" space, it's the spirit of experimentation. And with experimentation comes an inherent risk of failure. When setting out to build value-generating ML solutions, success is rarely guaranteed.
Adapting the famous adage: in data science project almost everything can go wrong, and so almost everything will go wrong.
Sometimes you can’t trust your data
Sometimes your model will just not have the predictive power
Sometimes you’ll find the business process cannot accommodate your brilliant technical solution.
Oh and sometimes, just as you’re putting buffed out solution in production, a market condition (looking at you, regulators) will change, render the whole solution useless and leaving to to start from scratch.
So how do we deal with this?..
The Zen of Failure in Data Science
I’ve found that seasoned data scientists often develop a zen-like attitude towards failure. They don't just accept failure; they expect it, prepare for it, embrace it when it happens, and - most importantly - learn from it. This mindset is so crucial that I’m convinced it's one of the defining features of a truly mature data scientist.
However, this skill isn't easy to acquire. As a data science leader, don't simply wait to see which team members naturally develop this ability. Instead, actively help your data scientists train and develop this skill, just as you would support them in mastering data manipulation or modelling techniques.
Here are some strategies to help your team embrace failure:
Hire for resilience: During the hiring process, thoroughly inquire about candidates' experiences with stress and setbacks. Make resilience an important factor in deciding who to hire.
Lead by example: When failure occurs, avoid pointing fingers. Instead, rally the team and focus on developing solutions together.
Implement “post-mortems”: Encourage your team to document why failures occurred and how they might be avoided in the future. Create a repo of failures and encourage everyone to consult it whenever necessary.
Celebrate learnings: Discuss insights gained from failures during team meetings. Praise team members who contribute valuable lessons learned from unsuccessful projects.
Balancing Failure and Value Delivery
When implementing the above, it is very important that you don’t accidentally glorify failure. That can definitely create an unhealthy fallback for when value is not delivered (“We botched a year-long strategic project, but hey - failure is okay!”) The key is to constantly stress that the good thing (or saving grace, really) about failing is only the useful learning that comes out of it. So, I would suggest stressing two points:
It’s okay to fail, but not okay to not learn from it.
If you have to fail (and as a data scientist, you do), fail early and fail often.
The second point is that while embracing failure is crucial, it's equally important to deliver value to stakeholders. Here's how to strike that balance:
Structure projects strategically: When starting a project, aim to deliver quick value before you go into experimentation which might yield nothing. This way you have some value to show if the experimentation does not pay off.
Contain the fallout: Do as small experiments as you can so that if they blow up, they don’t blow up the whole project.
Communicate effectively: If you’re going to experiment at any meaningful level, it falls on you as a leader to consult all stakeholders and have their buy-in in taking risks. This way you stress-test your ideas and have a shared responsibility about the outcome.
Measure and report: Develop metrics to quantify (even if not very precisely) the value of failed projects. For example:
Knowledge gained: e.g., number of new insights about data or business processes
Process improvements: e.g., identification of weak links in project process that leads to increase in future project success rate
Future cost savings: e.g., avoided expenses from identifying non-viable approaches early
Conclusion: Embracing Failure for Innovation
In the fast-paced world of data science, failure is not just inevitable—it's a crucial component of innovation and progress. By cultivating a team culture that views failure as a stepping stone to success, you'll be better equipped to tackle complex challenges, push boundaries, and deliver truly transformative solutions.
Remember, it's not about avoiding failure; it's about failing fast, learning faster, and continuously improving. The most innovative companies in the world, from tech giants to cutting-edge startups, have embraced this mindset. They understand that in the realm of data science, where we're often venturing into uncharted territory, failure is not just an acceptable outcome—it's often a necessary step on the path to groundbreaking discoveries.
As you lead your data science team, consider how you can create an environment where failure is not feared, but respected as a valuable teacher. How can you encourage your team to take calculated risks, to push the boundaries of what's possible, and to view setbacks as opportunities for growth?
How does your team handle project failures? Share your experiences and strategies in the comments below. Let's learn from each other and build more resilient, innovative data science teams together!