Skip to content

Teacher guide — Grade 10, Unit 5: What is a model? (and a first look at machine learning)

Time: 3 lessons + project (about 5 periods). Track: A then B.

# Lesson Time
1 Unplugged: models from cards 40–45 min
2 Code: k-NN and a baseline 40–45 min
3 Limits and misuse 40–45 min
P Unit project: Model vs baseline 2 periods

Before you start

  • Materials: Printed 30-card training set (visits, wealth → delivery in a facility); notebook notebooks/knn_intro.ipynb.
  • Read docs/pedagogy/lesson-anatomy.md and docs/pedagogy/ai-use-policy.md.

If you have no computers / no internet

The unplugged lesson is the heart of this unit: do k-NN on a printed grid with cards and coloured pins.

Common misconceptions

  • Models understand; they only find patterns in past data.
  • High accuracy on training data means a good model; test on unseen data.

Differentiation

  • Support: give a partially filled table / sentence starters; pair with a data-steward role.
  • Stretch: ask "what would change your conclusion?" and "who is missing?"

Answer key (check questions)

Lesson 1: Unplugged: models from cards 1. What is training data?
Answer: Examples the model learns from. 2. Why keep test data hidden?
Answer: To check it works on cases it has not seen. 3. Which cards are 'nearest'?
Answer: The ones with the closest values on the features.

Lesson 2: Code: k-NN and a baseline 1. What is a baseline?
Answer: A very simple rule to beat, e.g. always the majority. 2. Why test on unseen data?
Answer: Training accuracy can flatter the model. 3. k = 1 problem?
Answer: Follows noise; may overfit.

Lesson 3: Limits and misuse 1. Can a model be unfair?
Answer: Yes: it can copy bias in the training data. 2. Name a high-stakes use needing caution.
Answer: e.g. health care, loans, scholarships. 3. Who should decide?
Answer: People accountable for the outcome, not the model alone.

Assessment

Starter quiz each lesson (3 questions, one recalling the previous lesson); unit quiz in quiz.yml; project marked with the rubric in docs/pedagogy/assessment.md; 2-minute oral check per group.

Alignment

Optional Computer Science Grade 10: Python programming, databases and AI & contemporary technology; compulsory Maths data handling/statistics.

Answers: practice exercises

  1. (a) features humidity, pressure; label rain tomorrow (yes/no) → classification. (b) features area, location; label price → regression.
  2. Baseline 150/200 = 75%; model 160/200 = 80%: 5 percentage points better.
  3. √(3² + 4²) = 5.
  4. Distances from (2, 2): B 1.0, A 1.41, E 1.41, C 2.83, D 3.61. k = 1 → B → ✘. k = 3 → B ✘, A ✘, E ✔ → ✘.
  5. It is just a row number with no meaning; any "pattern" it finds is accidental (or reflects how the file was sorted), so it hurts predictions on new cases.
  6. The test set was used to choose k, so it is no longer unseen; 0.66 is optimistically biased. Split the training rows into train/validation, choose k on validation, then report test accuracy once.
  7. Training: baseline 0.558, rule 0.596; test: baseline 0.563, rule 0.660. The rule happens to do better on the test rows — with 300 cases, a few points of luck either way is normal.
  8. Unscaled age: k = 15 → 0.62, k = 25 → 0.593; age ÷ 5: both 0.62. Age adds little information here and, unscaled, dominates the distance; scaling reduces its influence.
  9. When the mistakes of the "more accurate" model fall on the most vulnerable group, or when the mistake it makes is the costly one (missing a high-risk mother), or when it can't be explained or audited.
  10. k = 15 accuracy by belt (test rows): Hill 0.620 (142), Mountain 0.636 (22), Terai 0.662 (136). Mountain is based on only 22 cases — very uncertain.

Real-data exercises: notes

  • R1 Bill length + flipper length separate the three species well: k-NN test accuracy typically ≈ 0.9+, baseline (Adelie) ≈ 0.44. Scaling matters (flipper mm range ~60, bill mm ~30).
  • R2 "Same as today" is a strong baseline for weather; in the monsoon, rain follows rain. A simple model may barely beat persistence — a good honest result.
  • R3 Census 2021: Mountain ≈ 6%, Hill ≈ 40%, Terai ≈ 54% of the population (check). The synthetic table has 10% Mountain, 46% Hill, 43% Terai — over-represents Mountain and Hill, under-represents Terai.