What Machine Learning Is

Rules you write, versus rules you find

Ordinary programming means writing the rule and letting the computer apply it:

def is_spam(subject):
    return "free money" in subject.lower()

print(is_spam("FREE MONEY now"))
print(is_spam("lunch tomorrow?"))
print(is_spam("f r e e   m o n e y"))

True
False
False

That third line is the problem. Every rule you add gets worked around, and the list of rules grows for ever.

Machine learning turns it around. You supply examples that are already labelled, and the program works out the rule:

examples in                        rule out
"FREE MONEY now"      -> spam
"lunch tomorrow?"     -> not spam        a model that scores
"f r e e m o n e y"   -> spam            any new subject line
...ten thousand more

Nobody writes the rule down. Nobody can fully read it afterwards either, which is the trade.

The words

The convention of capital X and lowercase y is everywhere in Python ML code: X is a table, y is a single column.

The three kinds

Supervised learning. You have the answers for your examples, and want them for new ones. Nearly all practical ML is this.

• Classification — the answer is a category: spam or not, which of three

species, which digit.

• Regression — the answer is a number: a price, a temperature, a duration.

Unsupervised learning. No answers. You are looking for structure — grouping customers who behave alike, or spotting the odd one out.

Reinforcement learning. The program acts, gets a score, and adjusts. Games and robotics, mostly, and a different toolkit.

answers = [("spam or not", "classification"),
           ("house price", "regression"),
           ("group similar users", "clustering")]

for question, kind in answers:
    print(f"{question:22} -> {kind}")

spam or not            -> classification
house price            -> regression
group similar users    -> clustering

It is arithmetic on numbers

A model has no idea what a flower or an email is. It sees a grid of numbers and finds arithmetic that maps that grid to the answers. Everything else — turning words into numbers, scaling, encoding categories — is preparation you do first.

That is also why data quality dominates. A model trained on wrong labels learns the wrong rule perfectly.

When not to use it

Reach for a rule you write yourself when:

• The rule is known. Tax is a formula. Do not train a model to add up.
• You have very little data. Under a few hundred examples, a person writing

rules usually wins.

• A wrong answer is expensive and you must explain it. "The model said so"

is not a reason anyone accepts for a refused loan.

• The pattern will not repeat. Models assume tomorrow resembles yesterday.

Use it when the rule is real but nobody can write it down: which photos contain a cat, which transactions are fraud, which sentence comes next.

What the rest of this section does

The next chapters build a working model end to end, then take it apart: how the data has to be shaped, why you hold some back, how to tell a good model from one that has merely memorised, and the mistakes that make a model look far better than it is.