Framing ML problems and the core vocabulary

Machine learning finds patterns in data to make predictions or decisions without hand-written rules. The first skill is not choosing an algorithm; it's framing the problem.

Types of learning

• Supervised: learn from labelled examples (input → known answer).

  • Regression: predict a number (salary, delivery time, demand).
  • Classification: predict a category (spam or not, will this candidate accept the offer).
• Unsupervised: find structure without labels: clustering (customer segments), dimensionality reduction (PCA), anomaly detection.
• Reinforcement learning: an agent learns by trial and reward (games, robotics, ad bidding).
• Self-supervised: the data provides its own labels (predict the next word); how large language models are pre-trained.

Framing checklist

1. What decision will the prediction drive, and what does a mistake cost?
2. What exactly is the target, and is it available at prediction time? (No future information, or "leakage".)
3. What data exists, how much, how is it labelled, and is it biased?
4. What's the baseline? A simple rule ("predict last week's value", "predict the majority class") that the model must beat.
5. Does it need ML at all? If a few rules work, ship the rules.

Core vocabulary

• Features (X): input variables. Label/target (y): what you predict.
• Training set to fit, validation set to tune, test set touched once at the end.
• Overfitting: memorises training noise; great on training data, poor on new data.
• Underfitting: too simple to capture the pattern; poor everywhere.
• Hyperparameters: settings you choose (tree depth, learning rate), unlike parameters the model learns (weights).
• Cross-validation: k-fold splits for a more reliable estimate with limited data.

The bias-variance trade-off

• High bias: wrong assumptions, too simple (a straight line for a curve) → underfits.
• High variance: too sensitive to the training data → overfits.
• Total error ≈ bias² + variance + irreducible noise. Increasing model complexity lowers bias but raises variance; the sweet spot is found with validation data.

Fixes for overfitting: more data, simpler model, regularisation (L1/L2), dropout, early stopping, fewer features, ensembling. Fixes for underfitting: more features, a more flexible model, less regularisation, training longer.

Data leakage, the silent killer

Leakage is when training uses information unavailable at prediction time: a "refund_issued" feature for predicting fraud, scaling with statistics from the whole dataset before splitting, or random splits on time-series data. The model looks brilliant offline and fails in production. Split first, fit preprocessing on training data only, and split time series by time.

Check yourself

• Classify these as regression, classification or clustering: predicting salary; grouping students by interests; detecting fake job posts.
• Explain overfitting to a non-technical manager.
• Give an example of data leakage.