Train the classifier with examples, not rules.
A 10-minute lesson: how machine learning feeds on examples, how signals and labels work, and a Trainer workshop to label real emails.
How machine learning works, what you teach it with examples and signals, and why labels are everything.
The case
A classifier doesn't understand your inbox: it learns from the examples you label. Label well, and it decides well.
In 10 minutes you'll train an email classifier: what it feeds on, how it sums signals, and why some examples are worth more than a thousand hand-written rules.
What you will learn
- Explain what a classifier learns during training
- Tell examples, signals, and labels apart
- Spot the trap of more data with bad labels
The trainer's recipe
Examples + signals + label. With those three, a classifier generalizes and nails what it never saw.
Examples
Solved cases that teach.
Signals
The traits it learns to look at.
Label
The decision that comes out when classifying.
Trap
More data doesn't fix loose labels.
Frequently asked questions
Does the classifier understand what an email means?
No: it sums signals and picks the label above the threshold. It compares patterns; it doesn't understand.
Are more training examples always better?
Only if the labels are right: more mislabeled data amplifies the error with confidence.
Test yourself: Lightning test: label or guess?
What does a classifier learn during training?
Correct answer: To look at signals from already-labeled examples.
With clean labels it learns which signals matter and generalizes: it nails cases it never saw.
You train with two emails and it fails on everything new. What's missing?
Correct answer: Examples: it has too little material to learn from.
Two examples don't reveal which signals matter. With more varied, well-labeled cases, it generalizes better.
You add 100,000 mislabeled emails. What happens?
Correct answer: It amplifies the error: it learns wrong signals with confidence.
Data doesn't fix loose labels: more trained garbage makes a classifier very confident about being wrong.
The pocket rule — Examples (what you show it) + signals (what it pays attention to) + label (what it decides). More data doesn't fix a loose label.
What you take away
Train a classifier with labeled examples and detect when data or signals fail.