The big idea: Classification predicts which category something belongs to, from labelled examples.
Spam or not. Benign or malignant. Which of ten digits. The output is a label, not a number on a scale.
| Binary | Multi-class | |
|---|---|---|
| Categories | Two | Three or more |
| Example | Spam / not spam | Which digit, 0–9 |
| Output | One probability | A probability per class |
| Decision | Compare with a threshold | Take the highest |
Free preview
This is the free notes preview
You're reading the free notes. Aimnova Pro unlocks the full study experience — and you can try it with your first topic free to keep:
- FlashcardsLock in vocabulary and key terms with spaced repetition.
- Practice questionsAnswer exam-style questions and get instant AI marking.
- Mock exams & past-paper vaultSit full mocks and see exactly how examiners award marks.
- Personalised study planA daily plan built around your exam date and weak areas.
Decision tree
- A sequence of yes/no questions down to a leaf
- Fully interpretable — the path IS the explanation
- Overfits easily unless its depth is limited
k-nearest neighbours
- Look at the k closest examples and take the majority
- No training at all — it just stores the data
- Slow at prediction, and needs features on comparable scales
Logistic regression
- Outputs a probability between 0 and 1
- Threshold it to get a class — and the threshold is a choice
- Despite the name, it classifies
The threshold is a decision, not a fact: A model outputs 0.7. Calling that "positive" uses a threshold of 0.5 — but nothing forces that.
For cancer screening you would lower it, accepting more false alarms to miss fewer real cases. Moving the threshold trades false positives against false negatives, and that is a human judgement about which error is worse.
Never wonder what to study next
Get a personalized daily plan based on your exam date, progress, and weak areas. We'll tell you exactly what to review each day.
| Measure | Answers | Matters when |
|---|---|---|
| Accuracy | What fraction were right? | Classes are balanced — and rarely otherwise |
| Precision | Of those flagged, how many really were? | A false alarm is costly |
| Recall | Of the real cases, how many were caught? | A miss is costly |
| Confusion matrix | All four outcomes at once | Always — it is the honest picture |
Accuracy lies on imbalanced data: On a disease affecting 1 in 1,000, a model answering "no" every time is 99.9% accurate and has never found a single case.
Always ask what a majority-class guesser would score. If it is close to your model, accuracy is telling you nothing.
Name which error is worse, and why: Full marks need the domain: a false negative in cancer screening means a missed diagnosis; a false positive means anxiety and a further test.
Say which matters more here, and what you would do to the threshold because of it.
How this is tested — you must go beyond accuracy to the two kinds of error, and tie the threshold to the cost of each. It comes up two ways:
Paper 1 Section A
- Explain how classification works, 3-4 marks
- Identify a task as binary or multi-class
- State what precision or recall measures
Paper 1 Section B — case study
- Evaluate a classifier from given figures
- Justify moving a decision threshold
The classic trap: Quoting accuracy alone. On imbalanced data it is close to meaningless — always separate false positives from false negatives and say which is worse in this domain.
A decision boundary — and the bent one that scores 100% and generalises worse.
Interactive diagram
Explore the labelled diagram, charts and maps for this topic in full study mode.
A model screening 10,000 scans finds 82 of the 100 tumours present, and wrongly flags 400 of the 9,900 healthy scans. Evaluate its performance and suggest what should change.
Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.