An introduction to Machine Learning
What the domain covers
This domain tests the machine-learning workflow: how models learn from data, the three main learning paradigms, the meaning of features and labels, and how to evaluate a model without fooling yourself. Expect questions on overfitting, data splits and choosing supervised versus unsupervised learning.
Core concepts
- Traditional programming uses rules plus data to produce answers. Machine learning uses data plus answers to produce a model.
- Supervised learning uses labelled examples (features plus target label). Classification predicts categories; regression predicts numbers.
- Unsupervised learning finds structure in unlabelled data: clustering (for example k-means), association and dimensionality reduction.
- Reinforcement learning trains an agent by reward and penalty through trial and error; there is no labelled correct answer, only feedback.
- Data split: the training set fits parameters; the validation set tunes choices; the test set gives a final unbiased estimate. Never tune on test data.
- Overfitting means memorising training noise and generalising poorly. Underfitting means the model is too simple and misses the pattern. More data, regularisation or a simpler model can help.
- Evaluation: accuracy can mislead on imbalanced data; use precision, recall, F1, a confusion matrix or ROC-AUC. Regression uses MAE or RMSE.
Worked example
Predict loan default from historical records containing features such as income, debt ratio and repayment history, plus the label defaulted or not. This is supervised binary classification. Split into train, validation and test; fit on train, tune on validation, then report precision and recall on test. Reasoning: labels exist, so it is not unsupervised; the target is categorical, so it is classification, not regression. If 95% of loans do not default, 95% accuracy may hide poor default detection, so inspect recall and precision. Check bias by group and avoid using post-decision information as a feature.
Traps that cost marks
- Unsupervised learning has no labels; clustering is not classification.
- Reinforcement learning uses reward, not a labelled dataset.
- Using test data during model selection inflates performance.
- Correlation in features does not prove causation.
- More complex models are not automatically better; overfitting is a risk.