Strip the buzzwords from applied machine learning and most of it is two jobs.
Regression: predict a number. What will this house sell for? How many kilowatts will the grid need at 6pm? Output: a value on a sliding scale, where close counts. Being $5k off on a house price is a good miss.
Classification: predict a category. Spam or not? Tumor or benign? Which of 10,000 species is in this photo? Output: a verdict, where close doesn't obviously exist — the email is junk or it isn't.
The split sounds cosmetic. It isn't: it decides the loss function (squared error vs cross-entropy), how you measure success, and what kinds of mistakes are even possible.
The threshold is a policy decision
Because classifiers really output probabilities, someone must choose where the verdict flips. A fraud model flagging at 0.5 catches less fraud but bothers fewer customers; at 0.2 it catches more and annoys thousands. Neither setting is "more accurate" — they're different trade-offs between false alarms and missed cases. The math can't make that call. It's a question about which mistake your institution can better afford, answered by a human and laundered through a decimal.
Where you've already met them
- Your inbox runs classification (spam), your bank runs both (fraud: classification; credit limit: regression).
- Weather apps: regression for temperature, classification for rain/no-rain.
- And the strangest example: a language model is a classifier run in a loop — at every step it classifies the context into one of ~100,000 possible next tokens, picks, and repeats. The chatbot writing your email is classification at machine-gun pace.
What to read next
Both jobs need labeled answers to learn from. What can a model find when nobody labeled anything? Clustering — structure without a teacher.