Framing a problem as an ML task

Before a single line of code: is this even an ML problem?

ML suits you when: the rule is hard to write but examples exist (images, text, behaviour), the pattern is stable over time, and mistakes are acceptable. ML does not suit you when: a rule is enough (VAT = 25 %), there is almost no data, every mistake is unacceptable, or the decision has to be legally explainable.

The framing (write it down):

• Input: exactly what the model sees at the moment of the decision — not things that become known later.
• Output: a class? a number? a ranking? Which question is being answered?
• Success metric: one number tied to real usefulness — and the level that is good enough.
• Baseline: what does the simplest rule give? ML has to beat it.

An example — «reduce food waste in the school canteen»:

Question	How many portions are needed tomorrow?
Input	day of the week, the dish, the number of pupils present yesterday, the weather, holidays
Output	a number (regression)
Metric	the absolute error in portions; the target is < 15
Baseline	the same as the same weekday last week
Data	the kitchen's logs for 2 years — do they exist? in what format?
Risk	a new menu → the pattern breaks; too few portions is worse than too many

Make the same table for: «flag abusive comments», «recommend the next exercise». Which of them should not be ML?