<aside> 🗺️
Module 1 · Is This Even an ML Problem? · Decision Makers · 60-90 minutes
</aside>
Your VP walks into the room. "Churn is up 15%. We need a machine learning model to fix it."
Sounds reasonable. But here's what happens next in most companies: a data scientist disappears for three months, builds something impressive-looking, and then... nothing. The model sits unused. The churn rate doesn't budge. Everyone's frustrated.
Why? Because nobody stopped to ask the basic questions. What does "fix churn" actually mean? What decision will this model inform? What data do we have? What does success look like?
This module is about those questions. Before you build anything, before you write a single line of code, you need to frame the problem correctly. Get this wrong and no algorithm can save you. Get it right and sometimes you realize you don't even need ML.
The question we're answering: How do I know if this is an ML problem, and how do I frame it so it actually gets used?
Think of ML as hiring a new employee and training them by showing examples of past decisions.
"Here's a customer who churned. Here's one who didn't. Here's another churner. Here's a loyal customer. Now—look at this new customer. What do you think they'll do?"
The employee learns patterns from examples. They can't follow explicit rules you give them (that's traditional programming). They can't know things you don't show them. And if you show them misleading examples—say, labeled incorrectly, or with information that wouldn't exist in the real world—they'll learn the wrong patterns.
This is exactly what ML does. Show it examples, it learns patterns, it makes predictions on new cases.
So the question becomes: Can I show it good examples? And will its predictions lead to useful actions?
Before reaching for ML, ask these questions:
| Question | If yes... | Do I have fewer than ~100 labeled examples? | Start with rules. ML needs data to learn. |
|---|---|---|---|
| Does a simple rule already achieve 90%+ of the value? | Keep the rule. Complexity has costs. | Is the pattern fixed and won't change? | Write the rule explicitly. ML is for patterns that drift. |
| Do I need 100% accuracy? | ML won't give you that. Consider rule-based systems with human review. |
The heuristic: Start with the simplest thing that works. Add ML when rules hit their ceiling and you have data to learn from.