Labels

A label is a category your function can predict. Labels define the decision space — the set of possible outputs your application will receive.

Good labels are:

• Mutually exclusive — each input clearly belongs to one label, not several
• Exhaustive — every input your function will see fits into one of your labels
• Concrete — named after real decisions your application needs to make

Good: Spam / Not spam
Problematic: Negative / Somewhat negative / Very negative (overlapping)

If two labels could reasonably apply to the same input, the model will struggle to learn the boundary between them.