Business requirements

Machine Learning for Business

Karolis Urbonas

Head of Machine Learning & Science, Amazon

Scoping business needs

What is the business situation?
- The company plans to expand to new markets
What is the business opportunity and how big is it?
- Identify the right markets with the biggest demand
What are the business actions we will take?
- Prioritize and invest more in the markets with higher predicted demand

Situation - The fraud rate has started increasing
Opportunity - Reduce fraud rate by X %, resulting in Y USD savings
Action - Work on improving fraud detection system, reduce fraud drivers, and manually review transactions at risk

fraud

Situation - The customers started to churn more
Opportunity - Reduce churn rate by X %, resulting in Y USD revenue saved
Action - Work on identifying and improving churn drivers (website errors, too much/little advertising, customer service issues etc.); identify customers at risk and introduce retention campaigns

churn

Always start with inference questions

Why has churn started increasing?

Which information indicates a potential transaction fraud?

How are our most valuable customers different from others?

Build on inference question to define prediction questions

Can we identify customers at risk of churning?

Can we flag potentially risky transactions?

Can we predict early on which customers are likely to become highly valuable?

Would you spend 1 million USD to earn extra 5000 USD each year? (~200 year return on investment)

Size up the opportunity
Once you know the drivers of the outcome, how much will it cost changing them, and what will be the value of doing that?
Finally, how do you know if you can affect the predicted outcome? (hint - experiments, experiments, and more experiments)

Finally, how do you know if you can affect the predicted outcome? (hint - experiments, experiments, and more experiments)

First, look at historical levels (churn, fraud, # of high value customers)
Run experiments e.g. target customers at risk with a discount, manually review top 10% riskiest transactions. Repeat experiments multiple times, see if you get a repeated pattern of desired results
If yes, use that to calculate opportunity and make decision if it's a worthwhile investment
If no - 1) collect more data, 2) qualitative research, 3) narrow down business question

Machine Learning for Business