Key Idea: The approaches divide by what feedback the learner gets: the correct answers (supervised), no answers at all (unsupervised), or a score for how things turned out (reinforcement). That one question places almost every method in this topic.
Paper 1
- Define a method; compute a measure.
- Identify which approach a scenario needs.
Paper 2
- Evaluate a model against an alternative.
- Reason about a reward or fitness function.
Both
- Name what is given up, not only what is gained.
🧭 Placing a method
| Feedback | Family | Methods here |
|---|---|---|
| The right answers | Supervised | Linear regression · classification |
| None | Unsupervised | Clustering · association rules |
| A reward | Reinforcement | Agents learning from consequences |
| A fitness score | Search, not learning | Genetic algorithms |
"How much?" is regression — a number on a scale. "Which one?" is classification — a category. The same data supports either; the question decides.
📏 Judging a model
Never quote accuracy alone
- Ask what a majority-class guesser would score — if it is close, accuracy is meaningless
- Precision — of those flagged, how many really were?
- Recall — of the real cases, how many were caught?
- Moving the threshold trades one against the other; it is a human judgement
- Cancer screening fears a false negative; a spam filter fears a false positive
Important: An association rule with 80% confidence is worthless if the consequent appears in 90% of baskets anyway — lift 0.89, meaning less than chance. Always divide by how common the consequent already is.
🧠 Networks
What the marks are for
- A unit is inputs × weights, summed, plus a bias, through an activation
- Without a non-linear activation, stacked layers collapse into one
- Training changes weights and biases; the structure is fixed beforehand
- A CNN adds convolution, shared weights and pooling — a feature is recognised wherever it appears
- Neither can explain itself, which is why interpretable models still win in regulated domains
⚖️ Choosing what to deploy
| Constraint | Choose | Because |
|---|---|---|
| A decision must be justified | Interpretable | A network has no rule to read out |
| Must run on a phone | Small | A large model cannot run there at all |
| Few hundred records | Simple | A complex model will fit noise |
| Nothing else constrained | Most accurate | Then accuracy is the criterion |
📝 Exam-style questions
Identify the machine learning approach for each, with a reason: (a) predicting next month's electricity demand, (b) grouping news articles with no categories defined, (c) teaching a drone to land, (d) deciding whether a transaction is fraudulent.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
A screening model catches 88 of 100 disease cases and wrongly flags 300 of 9,900 healthy people. A simpler, interpretable model catches 80 and wrongly flags 150. Evaluate which should be deployed.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
✅ Quick check
Cover the answers.
Which question places a method? What feedback does the learner get? Correct answers, none, or a reward.
Precision or recall — of the real cases, how many were caught? Recall. Precision is: of those flagged, how many really were?
Why is confidence alone misleading for association rules? It rewards whatever is popular. Lift divides by how common the consequent already is.
What does training a neural network change? The weights and biases only. Layers and units are hyperparameters, fixed beforehand.
When is the most accurate model the wrong choice? When a decision must be justified or audited, or when it cannot run where it is needed.
Exam tips
- Ask what feedback the learner gets — that places almost every method.
- "How much?" regression · "which one?" classification.
- Never quote accuracy alone; say what a majority-class guesser would score.
- Name which error the DOMAIN cannot afford, then move the threshold accordingly.
- Check lift, not just confidence.
- Say what you are giving up. A verdict with a condition attached scores highest.