Machine Learning Paradigms: Supervised Learning — Beginner-Friendly Notes

1. What Supervised Means (one line)

Learn a mapping input $x$ → output $y$ from labelled examples $\{(x_i, y_i)\}$ where the teacher already gives correct $y$. Goal: predict $y$ for unseen $x$.

If data comes with answers, it's supervised.

2. Two Types Inside Supervised

Type	Output	Example	Typical Metrics
Classification	discrete label (category)	spam/ham, cat/dog, disease yes/no	accuracy, precision, recall, F1
Regression	continuous number	house price, temperature, marks	MSE, RMSE, MAE

Beginner rule: label is word → classification; label is number → regression.

3. How It Works (conceptual pipeline)

Data collection → Cleaning → Train/Test split → Model choice → Training (minimise loss) → Evaluation → Deployment

• Training: adjust model parameters to minimise loss (error between $y$ and $\hat{y}$) on training set.
• Generalisation: must work on new data — not just memorise (overfitting).

4. Workflow with Train/Test Split

• Training set: learn from it.
• Validation set: tune hyperparameters, pick model.
• Test set: final unbiased evaluation — never touched during training.
• Overfitting: low train error, high test error → model memorised. Underfitting: high both → too simple. Fix with more data, simpler model, regularization.

5. Simple Examples

• Classification: email → spam/ham using words (Naive Bayes, decision tree, k-NN — next topics).
• Regression: house features (area, rooms) → price using linear regression (next).

6. Why Supervised Powers AI Today

Most deployed AI is supervised (vision, translation, fraud) because labelled data is abundant (web, phones) and evaluation is clear.

7. One-Line Summary

Supervised learning learns input-to-output maps from labelled examples for classification (discrete $y$) or regression (continuous $y$) via loss minimisation with train/validation/test evaluation to ensure generalisation.

8. Quick Self-Check

1. Is predicting student's grade A/B/C supervised? What type?
2. Why hold out a test set?