Data Science: How to frame the business problems as machine-learning problems (Part III)

In the previous articles (Part 1 and Part 2), we discussed when to use machine learning to solve a business problem and how to frame the business problems as a machine learning problem.

Now that we have developed a model, are we done? No. framing a business problem into a machine learning problem and solving the problem is only a part. The last mile is to convey the solution to the business users.

If you have been involving them at the discovery stage then the task is much simpler. But if not then the chances of the project failing despite what the data suggested is really high. Worse the teams keep struggling with the adoption problem. So, here are a few tips from my own experience on how to involve your business stakeholders in this journey and how to handover the model results to them:

Get early feedback on model quality: This follows from the suggestion above, do not surprise your business stakeholders with a model but get their feedback as often as possible, before you start working, iteratively as you are working on it, while you are running the test, before launching the results, and once you launch the results.

Embrace the fact that the Model performance does not translate one-on-one to business performance: Suppose the team worked on a recommender engine and test results show a 600% uplift in conversion. As tempting as it may be to announce that the model will lead to a 600% gain in conversion as evidenced from the test error — do not confuse the model performance with actual business performance. Setting such a high expectation will lead to disappointment even if the model gains a 25% uplift, which in all likelihood is a much more likely scenario given the tradeoff of many other dynamics that play around the actual product and user interactions.

While the above is true for any project, it is even more significant in the projects that involve machine learning.

Storytelling in simpler business words, not technical jargon: Saying we used the learning to rank method to solve the search relevancy issue, as tempting as it may be, does not help. You want your business stakeholders to collaborate with you not be amazed by your technical prowess. Using jargon, you risk alienating them, stop any further collaboration, so as much as possible de-jargonize your slides, and narrate the story from the problem space to the solution space, the model uplift and the expected business uplift in simpler easy to understand words of what you were trying to do, what methods you chose, what challenges you encountered, and what was the final conclusion. The goal at this stage should be to encourage discussions, feedback, and iteration ideas.

With this, I leave you to explore this beautiful world of translating the business problems into the data science/machine learning construct and vice versa to realize the true potential of collaboration between the business and data science team at your organization. If you have some feedback or comment, do reach out to me on Linkedin. I would love to stay in touch and learn from our collaboration. Alone, we can only go so far.

Read my other articles on Product Leadership, Product Growth, Pricing & Monetization strategy, and AI/ML here.