How does a spam filter learn? Teach one yourself

How does your email know that "WIN A FREE PRIZE NOW" is junk? Nobody sat down and wrote out every possible rule. Instead, a program looked at lots of examples and worked out the patterns by itself.

That's the big idea behind much of what people call AI: learning from examples instead of following written rules.

Teaching by example

Think of a young child learning what a dog is. You don't give them a definition. You point and say "dog" or "not a dog" until they get it.

A spam filter learns in a similar way, by counting. If "prize" keeps appearing in spam, and hardly ever in normal messages, it becomes a clue.

Try it: a tiny spam filter

You need to know no maths for this. Press Run, and see what the filter decides about the three new messages.

from collections import Counter

spam = ["win a free prize now", "free money click now", "claim your cash prize today"]
normal = ["see you at school tomorrow", "please bring your maths book", "the match is at five today"]

spam_words = Counter(" ".join(spam).split())
normal_words = Counter(" ".join(normal).split())

def check(message):
    spam_points = normal_points = 0
    for word in message.lower().split():
        spam_points += spam_words[word]
        normal_points += normal_words[word]
    answer = "SPAM" if spam_points > normal_points else "normal"
    print(f"{message!r}: spam clues {spam_points}, normal clues {normal_points} -> {answer}")

check("free cash prize")
check("bring your book to school")
check("win a maths prize")

Now you try

Press Edit, then try these. Before each run, guess what will happen.

1. Add your own test: add a line such as check("free school book").
2. Fool the filter: write a spammy message that it calls normal. Why did it work?
3. Teach it more: add "you won a lucky draw" to the spam list, then test check("lucky draw").

What you've just discovered

• It learned from data. Change the examples, and its behaviour changes, though you never touched the rules.
• More examples help. Our filter knows only a few dozen words. Real ones learn from huge numbers of messages.
• It can be wrong. It doesn't understand meaning, only patterns. That's true of bigger AI systems as well, so it's always worth checking what they tell you.
• Bad examples teach bad habits. If the examples are unfair or wrong, so is what it learns.

Where to go next

If you're curious how this works in real systems, the naive Bayes topic in our Machine Learning Fundamentals course builds a proper version. It's written for college students, but you may enjoy having a look.